brain-memory 0.1.0-beta.9 → 0.1.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +555 -0
- package/README.md +385 -84
- package/SECURITY.md +100 -0
- package/assets/brand/README.md +28 -0
- package/assets/brand/icon-dark.svg +19 -0
- package/assets/brand/icon-light.svg +20 -0
- package/assets/brand/linkedin_cover.png +0 -0
- package/assets/brand/linkedin_cover_light.png +0 -0
- package/assets/brand/mark-dark.svg +19 -0
- package/assets/brand/mark-light.svg +20 -0
- package/assets/brand/png/icon-dark-1024.png +0 -0
- package/assets/brand/png/icon-dark-128.png +0 -0
- package/assets/brand/png/icon-dark-256.png +0 -0
- package/assets/brand/png/icon-dark-512.png +0 -0
- package/assets/brand/png/icon-dark-64.png +0 -0
- package/assets/brand/png/icon-light-1024.png +0 -0
- package/assets/brand/png/icon-light-128.png +0 -0
- package/assets/brand/png/icon-light-256.png +0 -0
- package/assets/brand/png/icon-light-512.png +0 -0
- package/assets/brand/png/icon-light-64.png +0 -0
- package/assets/brand/png/mark-dark-1024.png +0 -0
- package/assets/brand/png/mark-light-1024.png +0 -0
- package/assets/icon.png +0 -0
- package/assets/icon.svg +19 -0
- package/bin/audit.js +93 -0
- package/bin/brain.js +132 -0
- package/bin/cloud-sync.js +150 -0
- package/bin/forget.js +163 -0
- package/bin/import.js +132 -0
- package/bin/install.js +63 -31
- package/bin/memorize.js +637 -0
- package/bin/pin.js +44 -0
- package/bin/recall.js +287 -0
- package/bin/reinforce.js +191 -0
- package/bin/restore.js +122 -0
- package/bin/session-start.js +337 -0
- package/bin/skill.js +67 -0
- package/bin/unpin.js +38 -0
- package/bin/verify.js +154 -0
- package/commands/brain/forget.md +243 -13
- package/commands/brain/import.md +125 -0
- package/commands/brain/memorize.md +95 -137
- package/commands/brain/pin.md +79 -0
- package/commands/brain/remember.md +53 -103
- package/commands/brain/skills.md +60 -0
- package/commands/brain/sleep.md +93 -6
- package/commands/brain/status.md +15 -6
- package/commands/brain/sync.md +181 -17
- package/commands/brain/verify.md +57 -0
- package/integrations/README.md +29 -0
- package/integrations/agentskills/brain-memory/SKILL.md +115 -0
- package/integrations/agentskills/brain-memory/scripts/session-start.mjs +34 -0
- package/package.json +19 -7
- package/prompts/antigravity.md +201 -0
- package/prompts/claude.md +107 -40
- package/prompts/copilot.md +203 -0
- package/prompts/kilo.md +202 -0
- package/prompts/openai.md +112 -45
- package/prompts/opencode.md +200 -0
- package/src/anomaly.js +260 -0
- package/src/audit.js +71 -0
- package/src/cloud-sync.js +672 -0
- package/src/content-lint.js +118 -0
- package/src/export-import.js +45 -7
- package/src/git-sync.js +170 -11
- package/src/harvest.js +562 -0
- package/src/index-manager.js +250 -11
- package/src/installer.js +250 -29
- package/src/pinning.js +146 -0
- package/src/provenance.js +79 -0
- package/src/quarantine.js +111 -0
- package/src/receipt.js +81 -0
- package/src/scorer.js +69 -12
- package/src/skills.js +208 -0
- package/src/tfidf.js +520 -0
- package/commands/brain/consolidate.md +0 -153
- package/commands/brain/explore.md +0 -79
- package/commands/brain/init.md +0 -149
- package/commands/brain/review.md +0 -130
- package/commands/brain/sunshine.md +0 -337
- package/prompts/gemini.md +0 -134
package/README.md
CHANGED
|
@@ -5,14 +5,14 @@
|
|
|
5
5
|
<h1 align="center">Brain Memory</h1>
|
|
6
6
|
|
|
7
7
|
<p align="center">
|
|
8
|
-
<a href="https://github.com/
|
|
8
|
+
<a href="https://github.com/omelas-tech/brain/actions/workflows/ci.yml"><img src="https://github.com/omelas-tech/brain/actions/workflows/ci.yml/badge.svg" alt="CI"></a>
|
|
9
9
|
<a href="https://www.npmjs.com/package/brain-memory"><img src="https://img.shields.io/npm/v/brain-memory" alt="npm version"></a>
|
|
10
|
-
<a href="https://github.com/
|
|
10
|
+
<a href="https://github.com/omelas-tech/brain/blob/main/LICENSE"><img src="https://img.shields.io/npm/l/brain-memory" alt="license"></a>
|
|
11
11
|
</p>
|
|
12
12
|
|
|
13
13
|
A hierarchical, file-system-based memory plugin for AI coding agents. Inspired by human neuroscience — memories are organized into deep nested life-domain categories, connected via associative networks, strengthened through spaced recall, and naturally decay over time.
|
|
14
14
|
|
|
15
|
-
|
|
15
|
+
One hosted **MCP connector** reaches every major agent and chat app — **Claude Code**, **OpenAI Codex CLI**, **OpenCode**, **GitHub Copilot CLI**, **Kilo**, the **Claude.ai** apps, **ChatGPT**, **Google Antigravity** (Gemini CLI's successor), **OpenClaw** / **NVIDIA NemoClaw**, **Hermes Agent**, and **Goose**. Prefer a free, local-first install? The native plugin runs entirely from `~/.brain/` on Claude Code, Codex, OpenCode, Copilot CLI, Kilo, and Antigravity — and dedicated [native integrations](integrations/) plug the same brain into OpenClaw and Hermes as their memory engine.
|
|
16
16
|
|
|
17
17
|
```
|
|
18
18
|
~/.brain/
|
|
@@ -40,86 +40,229 @@ Existing AI memory solutions use flat databases with tag-based retrieval. Brain
|
|
|
40
40
|
- **Human-inspectable** — Browse your "brain" in any file explorer
|
|
41
41
|
- **Git-friendly** — Full version history of how memories evolve
|
|
42
42
|
- **Strength + decay** — Recalled memories get stronger, forgotten ones fade. Just like your brain
|
|
43
|
+
- **Recall receipts** — Memory that visibly fires: every answer shaped by a memory ends with a one-line attributable receipt (`◉ memory: "<title>" (<type>, <age>)`), minted by the engine so it can't be hallucinated
|
|
44
|
+
- **Provenance-bounded trust** — Every memory records where it came from (`user`, `agent-inferred`, `tool-output`, `external`), and recall weighs it: a fact planted by a web page or tool output is down-ranked, flagged `low_trust`, marked `⚠` on its receipt, and can never outrank what you said directly — no matter how many copies of it get written
|
|
43
45
|
- **Associative network** — Memories link to each other with weighted connections. Recalling one activates related ones automatically
|
|
44
46
|
- **Context-dependent recall** — Memories encoded in a similar context to the current session are scored higher
|
|
45
47
|
- **Spaced reinforcement** — Memories recalled after longer intervals get bigger boosts, cramming produces diminishing returns
|
|
46
48
|
- **Cognitive types** — Episodic, semantic, and procedural memories each decay differently, just like in the brain
|
|
49
|
+
- **Always-present knowledge** — Pin critical conventions and preferences so they load every session, bypassing recall and never decaying — the facts your agent must never miss
|
|
50
|
+
- **Procedural skills** — Reusable how-to workflows with progressive disclosure, learned automatically from repeated experience and exportable to your agent's native skills
|
|
47
51
|
- **On-demand depth** — Subcategories are created as needed, not pre-defined
|
|
48
52
|
- **Consolidation** — Weak related memories merge into stronger combined knowledge
|
|
49
53
|
- **Zero dependencies** — Pure file I/O, no databases, no servers, no embeddings required
|
|
50
54
|
|
|
51
55
|
## Install
|
|
52
56
|
|
|
57
|
+
There are two ways to give an agent its brain. They are complementary — most people end up with both:
|
|
58
|
+
|
|
59
|
+
| Path | Reaches | Storage | Cost | Best for |
|
|
60
|
+
|------|---------|---------|------|----------|
|
|
61
|
+
| **Hosted MCP connector** | Every MCP-capable host (Claude Code, Codex CLI, OpenCode, Copilot CLI, Kilo, Claude.ai apps, ChatGPT, Antigravity, OpenClaw, Hermes, Goose) | Brain Cloud (hosted) | account | One connector, everywhere — including web & mobile chat |
|
|
62
|
+
| **Local-first native plugin** | The current CLIs (Claude Code, Codex CLI, OpenCode, Copilot CLI, Kilo, Antigravity) | `~/.brain/` on your machine | free | Zero-config ambient memory, no account, files you own |
|
|
63
|
+
| **[Assistant integrations](integrations/)** | OpenClaw / NVIDIA NemoClaw (memory-slot plugin), Hermes Agent (memory provider) | `~/.brain/` on your machine | free | Your always-on personal assistant sharing one brain with your coding agents |
|
|
64
|
+
|
|
65
|
+
### 1. Universal path — the hosted MCP connector
|
|
66
|
+
|
|
67
|
+
Brain Cloud exposes a single remote MCP server (streamable-HTTP + OAuth) at **`https://mcp.brainmemory.ai/mcp`**. Add it once per host and that host can recall, memorize, pin, and check status — no local install, no per-project setup. It works everywhere MCP does, including the chat apps that can't run a native plugin.
|
|
68
|
+
|
|
53
69
|
```bash
|
|
54
|
-
|
|
70
|
+
# Claude Code
|
|
71
|
+
claude mcp add --transport http brain https://mcp.brainmemory.ai/mcp
|
|
72
|
+
|
|
73
|
+
# OpenAI Codex CLI
|
|
74
|
+
codex mcp add brain --url https://mcp.brainmemory.ai/mcp
|
|
55
75
|
```
|
|
56
76
|
|
|
57
|
-
|
|
77
|
+
<details>
|
|
78
|
+
<summary>Per-host setup for every other connector</summary>
|
|
79
|
+
|
|
80
|
+
- **Codex CLI (config file)** — add to `~/.codex/config.toml`:
|
|
81
|
+
```toml
|
|
82
|
+
[mcp_servers.brain]
|
|
83
|
+
url = "https://mcp.brainmemory.ai/mcp"
|
|
84
|
+
```
|
|
85
|
+
- **OpenCode** — add to `opencode.json`:
|
|
86
|
+
```json
|
|
87
|
+
{
|
|
88
|
+
"mcp": {
|
|
89
|
+
"brain": { "type": "remote", "url": "https://mcp.brainmemory.ai/mcp" }
|
|
90
|
+
}
|
|
91
|
+
}
|
|
92
|
+
```
|
|
93
|
+
- **Claude.ai / ChatGPT** — **Settings → Connectors → Add custom connector**, paste `https://mcp.brainmemory.ai/mcp`, and complete the OAuth sign-in. (ChatGPT requires a paid plan or Developer Mode.)
|
|
94
|
+
- **Google Antigravity** — add to `mcp_config.json` under the `serverUrl` key. *(Legacy Gemini CLI used the `httpUrl` key in `settings.json`.)*
|
|
95
|
+
- **GitHub Copilot CLI** — add to `~/.copilot/mcp-config.json` (OAuth via `/mcp auth brain` in-session):
|
|
96
|
+
```json
|
|
97
|
+
{
|
|
98
|
+
"mcpServers": {
|
|
99
|
+
"brain-memory": { "type": "http", "url": "https://mcp.brainmemory.ai/mcp", "tools": ["*"] }
|
|
100
|
+
}
|
|
101
|
+
}
|
|
102
|
+
```
|
|
103
|
+
- **OpenClaw** — `openclaw mcp add brain-memory --url https://mcp.brainmemory.ai/mcp --auth oauth`, then `openclaw mcp login brain-memory`. (Inside NVIDIA NemoClaw, allowlist the endpoint first — see [integrations/openclaw](integrations/openclaw/).)
|
|
104
|
+
- **Hermes Agent** — add to `~/.hermes/config.yaml`:
|
|
105
|
+
```yaml
|
|
106
|
+
mcp_servers:
|
|
107
|
+
brain: { url: "https://mcp.brainmemory.ai/mcp", auth: oauth }
|
|
108
|
+
```
|
|
109
|
+
- **Goose** — add a remote extension (streamable-HTTP) pointing at `https://mcp.brainmemory.ai/mcp`.
|
|
110
|
+
- **Kilo** — add to `kilo.jsonc` (OAuth starts automatically):
|
|
111
|
+
```json
|
|
112
|
+
{
|
|
113
|
+
"mcp": {
|
|
114
|
+
"brain-memory": { "type": "remote", "url": "https://mcp.brainmemory.ai/mcp", "enabled": true }
|
|
115
|
+
}
|
|
116
|
+
}
|
|
117
|
+
```
|
|
118
|
+
|
|
119
|
+
</details>
|
|
120
|
+
|
|
121
|
+
The first time a host connects you'll complete an OAuth sign-in with the Google account that holds your cloud brain; the connector then syncs to every client on that account. See **[Use Brain in Claude apps](https://brainmemory.ai/docs/getting-started/claude-apps)** for the full walkthrough.
|
|
122
|
+
|
|
123
|
+
### 2. Local-first path — the native plugin
|
|
124
|
+
|
|
125
|
+
```bash
|
|
126
|
+
npm install -g brain-memory@beta
|
|
127
|
+
brain
|
|
128
|
+
```
|
|
129
|
+
|
|
130
|
+
The first command installs the package globally — this gives you both the interactive setup wizard and the `brain` CLI (`brain recall`, `brain memorize`, `brain reinforce`, …) that agents rely on for deterministic scoring.
|
|
131
|
+
|
|
132
|
+
The second command runs the setup wizard, which asks which runtime(s) to configure and whether to install globally or for the current project. Each runtime gets its instructions file plus the brain commands/skills:
|
|
133
|
+
|
|
134
|
+
| Runtime | Instructions file | Commands land at |
|
|
135
|
+
|---------|-------------------|------------------|
|
|
136
|
+
| **Claude Code** | `CLAUDE.md` | `~/.claude/commands/brain/` (plugin slash commands) |
|
|
137
|
+
| **OpenAI Codex CLI** | `AGENTS.md` (`~/.codex/`) | `~/.agents/skills/<name>/SKILL.md` |
|
|
138
|
+
| **OpenCode** | `AGENTS.md` | `~/.config/opencode/commands/` |
|
|
139
|
+
| **GitHub Copilot CLI** | `copilot-instructions.md` (`~/.copilot/`), repo-local `AGENTS.md` | `~/.agents/skills/<name>/SKILL.md` (shared with Codex) |
|
|
140
|
+
| **Kilo** | `rules/brain-memory.md` + `kilo.jsonc` `instructions` entry (`~/.config/kilo/`), repo-local `AGENTS.md` | `~/.config/kilo/commands/` |
|
|
141
|
+
| **Google Antigravity** | `GEMINI.md` | `~/.gemini/skills/` *(experimental — verify paths against a live install)* |
|
|
142
|
+
|
|
143
|
+
#### Non-interactive
|
|
144
|
+
|
|
145
|
+
```bash
|
|
146
|
+
brain --claude --global # Claude Code, global
|
|
147
|
+
brain --codex --global # OpenAI Codex CLI, global
|
|
148
|
+
brain --opencode --global # OpenCode, global
|
|
149
|
+
brain --copilot --global # GitHub Copilot CLI, global
|
|
150
|
+
brain --kilo --global # Kilo, global
|
|
151
|
+
brain --antigravity --global # Google Antigravity, global (experimental)
|
|
152
|
+
brain --all --global # All runtimes, global
|
|
153
|
+
```
|
|
154
|
+
|
|
155
|
+
For the deepest Copilot CLI integration — deterministic session-start context injection via its `sessionStart` hook — install the [Copilot plugin](integrations/copilot/) from the brain marketplace:
|
|
156
|
+
|
|
157
|
+
```bash
|
|
158
|
+
copilot plugin marketplace add omelas-tech/brain
|
|
159
|
+
copilot plugin install brain-memory@brain
|
|
160
|
+
```
|
|
58
161
|
|
|
59
|
-
|
|
162
|
+
Kilo gets the same treatment via its runtime plugin — deterministic session-start injection, session tracking, and `BRAIN_AGENT` labeling ([details](integrations/kilo/)):
|
|
60
163
|
|
|
61
164
|
```bash
|
|
62
|
-
|
|
63
|
-
|
|
64
|
-
npx brain-memory --codex --global # OpenAI Codex CLI, global
|
|
65
|
-
npx brain-memory --all --global # All runtimes, global
|
|
165
|
+
mkdir -p ~/.config/kilo/plugin
|
|
166
|
+
cp integrations/kilo/plugin/brain-memory.js ~/.config/kilo/plugin/
|
|
66
167
|
```
|
|
67
168
|
|
|
169
|
+
### 3. Personal-assistant integrations — OpenClaw, NemoClaw, Hermes
|
|
170
|
+
|
|
171
|
+
Always-on personal assistants get deeper, first-class integrations that make `~/.brain/` their actual memory engine — the same brain your coding agents read and write:
|
|
172
|
+
|
|
173
|
+
- **[OpenClaw / NVIDIA NemoClaw](integrations/openclaw/)** — a native **memory-slot plugin** (`openclaw plugins install openclaw-brain-memory` + `plugins.slots.memory: "brain-memory"`) that replaces `memory-core` with deterministic brain recall, session-start context injection, and model-driven capture. Also ships a slot-neutral hook pack, a ClawHub skill, and a NemoClaw egress-policy preset.
|
|
174
|
+
- **[Hermes Agent](integrations/hermes/)** — a native **memory provider** (`plugins/memory/brain/`) implementing the full `MemoryProvider` lifecycle: session-start prompt block, pre-turn prefetch, `brain_recall`/`brain_memorize`/`brain_reinforce` tools, and session-context logging. Plus lightweight hook-script glue if you keep another provider active.
|
|
175
|
+
|
|
176
|
+
Both are backed by the same `brain` CLI the coding-agent installs use — one memory, every agent.
|
|
177
|
+
|
|
68
178
|
### Update
|
|
69
179
|
|
|
70
180
|
```bash
|
|
71
|
-
|
|
181
|
+
npm install -g brain-memory@beta
|
|
182
|
+
brain update
|
|
72
183
|
```
|
|
73
184
|
|
|
74
|
-
|
|
185
|
+
The first command updates the package and the `brain` CLI. The second command refreshes the slash command prompts for your installed runtimes. Target specific runtimes with `--claude`, `--codex`, `--opencode`, `--antigravity`, or `--all`.
|
|
186
|
+
|
|
187
|
+
> **Upgrading from an older version?** The separate binaries (`brain-recall`, `brain-memorize`, `brain-reinforce`, `brain-cloud`, `brain-memory`) were unified into a single `brain` dispatcher — use `brain recall`, `brain memorize`, `brain reinforce`, `brain cloud <…>` instead. Run `brain update` to refresh your runtime prompts to the new command surface.
|
|
188
|
+
|
|
189
|
+
> **Why not `npx`?** `npx` runs the setup wizard in a temporary directory that is discarded after execution. The `brain` CLI (`brain recall`, `brain memorize`, `brain reinforce`) won't be available in your PATH, which means agents will fall back to less reliable manual file operations. Always use `npm install -g` to ensure everything works correctly.
|
|
75
190
|
|
|
76
191
|
### Uninstall
|
|
77
192
|
|
|
78
193
|
```bash
|
|
79
|
-
|
|
194
|
+
brain uninstall
|
|
195
|
+
npm uninstall -g brain-memory
|
|
80
196
|
```
|
|
81
197
|
|
|
82
|
-
|
|
198
|
+
The first command removes slash commands and prompt sections from your agent configs. The second removes the package and CLI tools. Your `~/.brain/` directory (memories) is preserved by default. Add `--delete-data` to the uninstall command to remove it too. Use `--yes` to skip confirmation prompts.
|
|
83
199
|
|
|
84
200
|
### Manual Install
|
|
85
201
|
|
|
86
|
-
|
|
202
|
+
If you'd rather wire up the native plugin by hand, copy the brain commands into your agent's folder, then append the matching prompt file to its instructions file.
|
|
87
203
|
|
|
88
204
|
```bash
|
|
89
|
-
# Claude Code
|
|
205
|
+
# Claude Code — flat slash commands
|
|
90
206
|
cp -r commands/brain/ ~/.claude/commands/brain/
|
|
91
207
|
|
|
92
|
-
#
|
|
93
|
-
|
|
208
|
+
# OpenCode — flat slash commands
|
|
209
|
+
mkdir -p ~/.config/opencode/commands/brain
|
|
210
|
+
cp -r commands/brain/. ~/.config/opencode/commands/brain/
|
|
94
211
|
|
|
95
|
-
# OpenAI Codex CLI
|
|
212
|
+
# OpenAI Codex CLI — each command becomes a skill under the cross-tool skills dir
|
|
96
213
|
for f in commands/brain/*.md; do
|
|
97
214
|
name=$(basename "$f" .md)
|
|
98
|
-
mkdir -p ~/.
|
|
99
|
-
cp "$f" ~/.
|
|
215
|
+
mkdir -p ~/.agents/skills/brain-"$name"
|
|
216
|
+
cp "$f" ~/.agents/skills/brain-"$name"/SKILL.md
|
|
217
|
+
done
|
|
218
|
+
|
|
219
|
+
# Google Antigravity (experimental — verify paths) — skills under ~/.gemini/skills/
|
|
220
|
+
for f in commands/brain/*.md; do
|
|
221
|
+
name=$(basename "$f" .md)
|
|
222
|
+
mkdir -p ~/.gemini/skills/brain-"$name"
|
|
223
|
+
cp "$f" ~/.gemini/skills/brain-"$name"/SKILL.md
|
|
100
224
|
done
|
|
101
225
|
```
|
|
102
226
|
|
|
103
227
|
Then append the contents of the corresponding prompt file to your agent's instructions file:
|
|
104
228
|
- `prompts/claude.md` → `CLAUDE.md`
|
|
105
|
-
- `prompts/
|
|
106
|
-
- `prompts/
|
|
229
|
+
- `prompts/openai.md` → `AGENTS.md` (Codex, in `~/.codex/`)
|
|
230
|
+
- `prompts/opencode.md` → `AGENTS.md` (OpenCode)
|
|
231
|
+
- `prompts/antigravity.md` → `GEMINI.md` (experimental)
|
|
232
|
+
|
|
233
|
+
> **`AGENTS.md` collision (local scope):** Codex and OpenCode both read a project-local `AGENTS.md`. If you install both for the *same* project, point only one at the local file (or install one globally) so their brain sections don't conflict. Globally they're separate (`~/.codex/AGENTS.md` vs OpenCode's own config dir).
|
|
107
234
|
|
|
108
235
|
## Commands
|
|
109
236
|
|
|
237
|
+
The everyday loop is **ambient** — `remember` and `memorize` run automatically at session start/end, so you rarely type a command. The surface is a small core plus a background maintenance job:
|
|
238
|
+
|
|
110
239
|
| Command | Description |
|
|
111
240
|
|---------|-------------|
|
|
112
|
-
| `/brain:init` | Initialize `~/.brain/` directory structure with default categories |
|
|
113
|
-
| `/brain:memorize [topic]` | Store memories from current session context |
|
|
114
241
|
| `/brain:remember [query]` | Recall relevant memories with spreading activation and context matching |
|
|
115
|
-
| `/brain:
|
|
116
|
-
| `/brain:explore [category]` | Browse the brain hierarchy with visual tree view |
|
|
117
|
-
| `/brain:consolidate [scope]` | Merge related weak memories into stronger combined ones |
|
|
118
|
-
| `/brain:forget [target]` | Decay, archive, or remove memories |
|
|
119
|
-
| `/brain:sunshine [target]` | Deep forensic erasure — trace and remove all references |
|
|
120
|
-
| `/brain:sleep [scope]` | Full maintenance cycle — 9 neuroscience-inspired phases |
|
|
242
|
+
| `/brain:memorize [topic] [--sync]` | Store memories from current session context (add `--sync` to auto-push) |
|
|
121
243
|
| `/brain:status` | Dashboard with brain health metrics and recommendations |
|
|
122
|
-
| `/brain:
|
|
244
|
+
| `/brain:pin [id\|query]` | Pin a memory to the always-present tier — loads every session, never decays. Toggles: also unpins (`--off`) |
|
|
245
|
+
| `/brain:forget [target]` | Decay, archive, or remove memories. `--deep` performs forensic erasure — traces and removes every reference |
|
|
246
|
+
| `/brain:import [--project P] [--since 30d]` | Cold-start a new brain from transcripts your agents already wrote. Incremental — safe to re-run |
|
|
247
|
+
| `/brain:sync [subcommand]` | Sync via Brain Cloud, Git remote, or export/import (auto-initializes the brain on first run) |
|
|
248
|
+
| `/brain:skills [list\|show\|add\|use\|remove\|export]` | Manage procedural skills — reusable how-to workflows with progressive disclosure |
|
|
249
|
+
| `/brain:sleep [scope]` | Full maintenance cycle — 9 neuroscience-inspired phases (replay, consolidation, review reinforcement, pruning, dreaming, …). Usually runs automatically/in the background |
|
|
250
|
+
|
|
251
|
+
## Cold Start
|
|
252
|
+
|
|
253
|
+
A new brain is empty, and an empty brain is worth nothing until something fills it. But your agents have been keeping local transcripts for months — decisions, preferences, corrections, conventions. `/brain:import` reads them.
|
|
254
|
+
|
|
255
|
+
```bash
|
|
256
|
+
brain import --sources # what history exists on this machine
|
|
257
|
+
brain import --project my-app # digest that project's past sessions
|
|
258
|
+
brain import --since 30d # or scope by time
|
|
259
|
+
```
|
|
260
|
+
|
|
261
|
+
The split matters. The CLI **harvests** — deterministic, budget-bounded, no model call: it extracts session titles, user prompts, projects, branches, and edited files, filters out harness noise and subagent traffic, and interleaves across projects so one busy repo can't eat the budget. The agent then **distills** that digest into memories through the ordinary `brain memorize` path.
|
|
262
|
+
|
|
263
|
+
That split is deliberate. Extracting meaning is a semantic judgement, and doing it in the CLI would mean shipping an embedding model or calling an LLM — the two things Brain exists to avoid. So the harvester reports facts and never infers, and every imported memory lands as `agent-inferred` provenance: it inherits the existing ceilings for free, and **an import can never pin, entrench, or outrank something you said directly.**
|
|
264
|
+
|
|
265
|
+
Import is incremental. A cursor in `~/.brain/.import/state.json` records which sessions have been read, so re-running only offers new ones (`--all` overrides). Nothing is written without you — `brain import` only ever reads and prints.
|
|
123
266
|
|
|
124
267
|
## Session Lifecycle
|
|
125
268
|
|
|
@@ -127,22 +270,41 @@ Brain Memory works automatically in the background — no commands needed for ba
|
|
|
127
270
|
|
|
128
271
|
### Session Start
|
|
129
272
|
|
|
130
|
-
When a session begins and `~/.brain/` exists, the agent
|
|
273
|
+
When a session begins and `~/.brain/` exists, the agent makes a single `brain session-start` call. This aggregator returns one deterministic, **token-budget-bounded** payload so the brain can never bloat the context window (the budget lives in `~/.brain/config.json`):
|
|
131
274
|
|
|
132
|
-
1.
|
|
133
|
-
2.
|
|
134
|
-
3.
|
|
275
|
+
1. **Pinned memories** — always-present conventions and preferences, injected verbatim regardless of recall score
|
|
276
|
+
2. **Skills index** — the name + description of each available procedural skill (~100 tokens each; full instructions are loaded only when a task matches)
|
|
277
|
+
3. **Context recall** — the top memories relevant to the current project
|
|
278
|
+
4. **Review queue + low-confidence alerts**, then a brief status line:
|
|
135
279
|
|
|
136
280
|
```
|
|
137
|
-
|
|
138
|
-
📋 3 memories due for review —
|
|
281
|
+
◉ Brain active — 42 memories loaded (8 in current project context)
|
|
282
|
+
📋 3 memories due for review — reinforced automatically during /brain:sleep
|
|
139
283
|
```
|
|
140
284
|
|
|
141
|
-
The agent silently internalizes relevant memories and references them naturally during the session — no information dump.
|
|
285
|
+
The agent treats pinned facts as active constraints, notes which skills exist, silently internalizes relevant memories, and references them naturally during the session — no information dump.
|
|
286
|
+
|
|
287
|
+
### Ambient Session Tracking
|
|
288
|
+
|
|
289
|
+
Throughout the session, the agent maintains a running mental log of notable events — decisions made, things learned, insights realized, significant experiences. This is purely internal awareness with no file writes. It ensures that early-session events aren't forgotten by the time the session ends.
|
|
290
|
+
|
|
291
|
+
### Periodic Memory Checkpoint
|
|
292
|
+
|
|
293
|
+
Every ~10 substantive interactions (file edits, architecture decisions, debugging breakthroughs), the agent evaluates whether memorizable content has accumulated. If so, it appends a brief reminder to its next response:
|
|
294
|
+
|
|
295
|
+
```
|
|
296
|
+
◉ Notable decisions and learnings this session — /brain:memorize when ready
|
|
297
|
+
```
|
|
298
|
+
|
|
299
|
+
This never interrupts your flow — it's a one-liner appended to an existing response, at most once per ~10 interactions. The counter resets after you run `/brain:memorize`.
|
|
142
300
|
|
|
143
301
|
### Session End
|
|
144
302
|
|
|
145
|
-
When a session ends, the agent
|
|
303
|
+
When a session ends, the agent performs two steps in order:
|
|
304
|
+
|
|
305
|
+
1. **Saves session context** to `~/.brain/contexts.json` — always, even for trivial sessions. This includes a `notable_unsaved` field listing items you didn't memorize, so future sessions can reference what happened.
|
|
306
|
+
|
|
307
|
+
2. **Suggests memorization** if the session contained meaningful decisions, learnings, or insights:
|
|
146
308
|
|
|
147
309
|
```
|
|
148
310
|
💡 This session contained notable decisions and learnings.
|
|
@@ -150,7 +312,7 @@ When a session ends, the agent evaluates whether significant decisions, learning
|
|
|
150
312
|
Run /brain:memorize to capture them before this context is lost.
|
|
151
313
|
```
|
|
152
314
|
|
|
153
|
-
The agent never auto-memorizes without user consent.
|
|
315
|
+
The agent never auto-memorizes without user consent. Context is saved proactively — the agent doesn't wait for an explicit goodbye signal.
|
|
154
316
|
|
|
155
317
|
## How It Works
|
|
156
318
|
|
|
@@ -167,8 +329,8 @@ Create → Store → Decay → Recall → Reinforce → Review → Sleep → Arc
|
|
|
167
329
|
3. **Decay** — Memories naturally weaken over time: `effective_strength = base_strength * (decay_rate ^ days_since_access)`
|
|
168
330
|
4. **Recall** — `/brain:remember` searches, scores with spreading activation and context matching, and returns the best memories. Archived memories are searched as a fallback.
|
|
169
331
|
5. **Reinforce** — Each recall applies spaced reinforcement (longer gaps = bigger boosts) and improves the memory's decay resistance
|
|
170
|
-
6. **Review** —
|
|
171
|
-
7. **Sleep** — `/brain:sleep` performs a 9-phase maintenance cycle inspired by real neuroscience
|
|
332
|
+
6. **Review** — SM-2 spaced repetition surfaces memories at optimal intervals for long-term retention; reinforcement runs automatically during the sleep cycle
|
|
333
|
+
7. **Sleep** — `/brain:sleep` performs a 9-phase maintenance cycle inspired by real neuroscience (replay, consolidation, review reinforcement, pruning, dreaming, …), usually run automatically/in the background
|
|
172
334
|
8. **Archive** — Fully decayed memories move to `_archived/` (recoverable and searchable) or can be permanently deleted
|
|
173
335
|
|
|
174
336
|
### Neuroscience Foundations
|
|
@@ -214,6 +376,38 @@ Each memory is also classified by how the brain processes it:
|
|
|
214
376
|
|
|
215
377
|
During sleep, frequently-recalled episodic memories are **crystallized** into semantic memories — the specific event fades but the lesson persists. This mirrors how humans extract general principles from repeated experiences.
|
|
216
378
|
|
|
379
|
+
### Always-Present Memories (Pinning)
|
|
380
|
+
|
|
381
|
+
Recall is probabilistic — a stored preference only applies if scoring happens to surface it. For the facts an agent must **never** miss (coding conventions, standing decisions, hard constraints), that's not good enough. Pinning fixes it:
|
|
382
|
+
|
|
383
|
+
- **`pinned`** — the memory is injected at **every** session start regardless of recall score, and is **decay-exempt** (it never fades). Scope it `global` (loads everywhere) or `project:<name>` (loads only in that project).
|
|
384
|
+
- **`stable`** — an independent flag that exempts a memory from decay and pruning **without** forcing it to always load — for timeless facts you recall on demand but don't want to fade.
|
|
385
|
+
|
|
386
|
+
```bash
|
|
387
|
+
brain pin <id> --scope global --priority 1 # or: /brain:pin always use tabs
|
|
388
|
+
```
|
|
389
|
+
|
|
390
|
+
Pinned and stable memories are skipped by sleep-cycle homeostasis and pruning, and `/brain:memorize` proactively proposes pinning durable conventions. The always-present set is budget-capped (`pin_budget_tokens`) so it can't crowd out the context window. This maps cleanly to the **semantic memory** type in the [CoALA](https://arxiv.org/abs/2309.02427) agent-memory model — knowledge that is always in context — decomposed into two orthogonal properties (always-loaded vs. non-decaying).
|
|
391
|
+
|
|
392
|
+
### Procedural Skills
|
|
393
|
+
|
|
394
|
+
Procedural memory is *how* to do things — reusable, step-by-step workflows stored as `~/.brain/_skills/<name>/SKILL.md`. To keep them from flooding the context window, skills use **three-level progressive disclosure**:
|
|
395
|
+
|
|
396
|
+
| Level | When | What loads |
|
|
397
|
+
|-------|------|------------|
|
|
398
|
+
| **L0** | Every session start | Only each skill's name + description (~100 tokens) |
|
|
399
|
+
| **L1** | A task matches a skill | The full `SKILL.md` step-by-step instructions |
|
|
400
|
+
| **L2** | A step needs them | Referenced `resources/` (templates, scripts) |
|
|
401
|
+
|
|
402
|
+
```bash
|
|
403
|
+
brain skill list # advertised skills (L0)
|
|
404
|
+
brain skill show structured-code-review # full instructions (L1)
|
|
405
|
+
brain skill use structured-code-review # record outcome (--failed weakens it)
|
|
406
|
+
brain skill export structured-code-review --target claude # → native .claude/skills/
|
|
407
|
+
```
|
|
408
|
+
|
|
409
|
+
Skills strengthen on successful `use` and weaken on `--failed` — one that fails too often demotes itself out of the advertised L0 index. And they aren't only hand-authored: during sleep, Brain **crystallizes** recurring task-solving patterns from repeated experience into new skills (user-confirmed). Finally, `brain skill export` emits a skill in your agent's **native** format so it becomes directly executable.
|
|
410
|
+
|
|
217
411
|
### Memory File Format
|
|
218
412
|
|
|
219
413
|
Each memory is a Markdown file with YAML frontmatter:
|
|
@@ -282,6 +476,16 @@ score = 0.38 * relevance
|
|
|
282
476
|
- **context_match** (0.14) — How similar the encoding context is to the current session
|
|
283
477
|
- **salience** (0.08) — Emotional/motivational significance
|
|
284
478
|
|
|
479
|
+
Relevance is **calibrated in absolute terms**: BM25 scores are scaled by
|
|
480
|
+
IDF-weighted *query coverage*, so a memory matching one term of a four-term query
|
|
481
|
+
cannot claim relevance 1.0 just for being the best available match. Explicit-query
|
|
482
|
+
recall also applies a **relevance floor** — memories that are neither
|
|
483
|
+
query-relevant nor activated by a relevant associate are excluded rather than
|
|
484
|
+
padding the results on strength alone. A query about things the brain doesn't
|
|
485
|
+
know returns few results with honestly low scores (or none), instead of
|
|
486
|
+
confident-looking noise. Context-mode recall (session start) is exempt, where
|
|
487
|
+
strength-ranked topical padding is the intended behavior.
|
|
488
|
+
|
|
285
489
|
The agent then decides the response strategy:
|
|
286
490
|
- **Single strong match** (top score > 0.7) → Return the full memory
|
|
287
491
|
- **Multiple related** (2-5 candidates > 0.4) → Synthesize a consolidated response
|
|
@@ -374,7 +578,7 @@ Each expertise area gets an `_expertise.md` profile documenting what you know we
|
|
|
374
578
|
|
|
375
579
|
### Spaced Repetition Review
|
|
376
580
|
|
|
377
|
-
|
|
581
|
+
Brain Memory implements the SM-2 algorithm to surface memories at optimal intervals for long-term retention. The review queue is generated and reinforced during `/brain:sleep`, and tracks:
|
|
378
582
|
|
|
379
583
|
- **Interval** — Time until next review (grows exponentially with successful recalls)
|
|
380
584
|
- **Ease factor** — How easily the memory is recalled (adjusts based on recall quality 1-5)
|
|
@@ -384,17 +588,27 @@ Failed recalls reset the interval to 1 day. Successful recalls extend the interv
|
|
|
384
588
|
|
|
385
589
|
### Cross-Agent Memory Sharing
|
|
386
590
|
|
|
387
|
-
`~/.brain/` is a single global directory in the user's home folder. All memories are shared across every project and every supported agent automatically. A decision stored by Claude Code in one project is immediately available to
|
|
591
|
+
`~/.brain/` is a single global directory in the user's home folder. All memories are shared across every project and every supported agent automatically. A decision stored by Claude Code in one project is immediately available to OpenAI Codex CLI, OpenCode, or Google Antigravity in any other project — no configuration, no export, no per-project setup. The format is agent-agnostic: plain Markdown files with YAML frontmatter, readable by any tool.
|
|
592
|
+
|
|
593
|
+
Brain Memory is **model-agnostic** as well as agent-portable: because memory is plain files rather than embeddings welded to a particular model, the LLM underneath your agent can be anything — GPT, Claude, Gemini, or any model routed through a gateway — and your memory is unaffected. **Switch your model or switch your agent, and you keep remembering and pick up exactly where you left off.**
|
|
388
594
|
|
|
389
595
|
To share memories across different machines, use `/brain:sync` (see below).
|
|
390
596
|
|
|
391
|
-
### Portable Sync
|
|
597
|
+
### Portable Sync — your memory, no lock-in
|
|
598
|
+
|
|
599
|
+
**Your brain is plain files in a folder — sync it however you already sync files.** No account or cloud provider is required.
|
|
600
|
+
|
|
601
|
+
**Any synced folder** — Point `BRAIN_DIR` at a folder your existing tools already sync, and you're done:
|
|
392
602
|
|
|
393
|
-
|
|
603
|
+
```bash
|
|
604
|
+
export BRAIN_DIR="$HOME/Google Drive/brain" # or Dropbox, iCloud Drive, OneDrive, Syncthing…
|
|
605
|
+
```
|
|
394
606
|
|
|
395
|
-
**Git remote** —
|
|
607
|
+
**Git remote** — `/brain:sync` push/pull `~/.brain/` to any private Git repository (GitHub, GitLab, Codeberg, or self-hosted), using your existing Git/SSH auth.
|
|
396
608
|
|
|
397
|
-
**Export/Import** — Pack the entire
|
|
609
|
+
**Export/Import** — Pack the entire brain into a single encrypted file for manual transfer via USB, email, or any file-sharing service.
|
|
610
|
+
|
|
611
|
+
**Brain Cloud** (optional) — A hosted, zero-config hub if you'd rather not run your own. Entirely optional — everything above works without it.
|
|
398
612
|
|
|
399
613
|
**Key features:**
|
|
400
614
|
- **Manual push/pull** — No background watchers, no auto-sync. You control when data moves.
|
|
@@ -409,16 +623,52 @@ To share memories across different machines, use `/brain:sync` (see below).
|
|
|
409
623
|
|
|
410
624
|
For one-off transfers, use `/brain:sync export` and `/brain:sync import <path>`.
|
|
411
625
|
|
|
626
|
+
**Restore — an undo button for your whole brain.** Every push is a restore point, and the CLI can roll the entire brain back to any of them:
|
|
627
|
+
|
|
628
|
+
```bash
|
|
629
|
+
brain restore --list # list restore points (the Git sync history)
|
|
630
|
+
brain restore --to <commit> # roll ~/.brain/ back to that point
|
|
631
|
+
brain restore --list --from cloud # list Brain Cloud's pre-push snapshots
|
|
632
|
+
brain restore --to <version> --from cloud # restore from a cloud snapshot
|
|
633
|
+
```
|
|
634
|
+
|
|
635
|
+
Before touching anything, restore preserves your *current* state — the Git path commits it as a safety snapshot, the cloud path writes a local backup under `~/.brain/.cloud/` — so a restore is always undoable, even for work you never pushed. (Brain Cloud also snapshots the brain it's about to replace on every push, server-side.) The append-only `audit.log` is carried forward through restores (never rolled back), and every restore is itself logged there. Restore granularity follows push frequency: push often for finer-grained history.
|
|
636
|
+
|
|
412
637
|
Sync state is stored locally in `~/.brain/.sync/` and is never pushed to the remote.
|
|
413
638
|
|
|
639
|
+
### Brain in Claude apps (web & mobile) — beta
|
|
640
|
+
|
|
641
|
+
Beyond the CLI, you can use your brain from the **Claude apps** (web, desktop, and your
|
|
642
|
+
**phone**) via the **Brain connector** — a remote MCP server hosted by [Brain Cloud](https://app.brainmemory.ai).
|
|
643
|
+
Add it once on claude.ai (**Settings → Connectors → Add custom connector → `https://mcp.brainmemory.ai/mcp`**),
|
|
644
|
+
sign in with the Google account holding your cloud brain, and it syncs to every Claude client on
|
|
645
|
+
your account. Read **and write**: recall + status, plus memorize / pin / unpin (writes prompt for
|
|
646
|
+
confirmation) — so you can capture a memory on your phone and recall it on your laptop.
|
|
647
|
+
|
|
648
|
+
→ Full guide: **[Use Brain in Claude apps](https://brainmemory.ai/docs/getting-started/claude-apps)**
|
|
649
|
+
|
|
650
|
+
**Your data, protected.** Brains synced to Brain Cloud are **encrypted at rest**
|
|
651
|
+
(AES-256-GCM, per-user key); the connector holds working copies in **RAM only**
|
|
652
|
+
(never on disk); all traffic is **TLS**; sessions use **rotating, revocable tokens
|
|
653
|
+
with reuse detection** (log out one device or all); and every request is isolated
|
|
654
|
+
to your account. Prefer the server never sees plaintext? Stay **local-only** (the
|
|
655
|
+
default) or use Git/export sync with a passphrase. See **[SECURITY.md](SECURITY.md)**.
|
|
656
|
+
|
|
414
657
|
## File Structure
|
|
415
658
|
|
|
416
659
|
```
|
|
417
660
|
~/.brain/
|
|
418
661
|
├── index.json # Memory inventory — fast lookup for all memories
|
|
662
|
+
├── config.json # Working-memory token budgets (session-start injection)
|
|
419
663
|
├── associations.json # Weighted associative network between memories
|
|
420
664
|
├── contexts.json # Session context snapshots for context-dependent recall
|
|
421
665
|
├── review-queue.json # Spaced repetition scheduling
|
|
666
|
+
├── pinned.json # Always-present tier manifest (pinned memory IDs + scope)
|
|
667
|
+
├── skills-index.json # Advertised procedural skills (L0 — name + description)
|
|
668
|
+
├── _skills/ # Procedural skills
|
|
669
|
+
│ └── <skill-name>/
|
|
670
|
+
│ ├── SKILL.md # Advertised description + step-by-step instructions
|
|
671
|
+
│ └── resources/ # Optional templates/scripts (loaded only at execution)
|
|
422
672
|
├── professional/ # Work, career, technical skills
|
|
423
673
|
│ ├── _meta.json # Category metadata and stats
|
|
424
674
|
│ ├── _expertise.md # Generated expertise profile
|
|
@@ -489,35 +739,58 @@ Brain configuration lives in `~/.brain/index.json` under the `config` key:
|
|
|
489
739
|
| `association_config.spreading_activation_depth` | 2 | Maximum hops for spreading activation traversal |
|
|
490
740
|
| `association_config.spreading_activation_decay` | 0.5 | Decay factor per hop during spreading activation |
|
|
491
741
|
|
|
492
|
-
|
|
742
|
+
### Working-Memory Budget
|
|
743
|
+
|
|
744
|
+
A separate `~/.brain/config.json` (created lazily with safe defaults) caps how much the brain injects into the context window at session start, so the always-present tier can never crowd out your actual work:
|
|
745
|
+
|
|
746
|
+
```json
|
|
747
|
+
{
|
|
748
|
+
"working_memory_budget_tokens": 3000,
|
|
749
|
+
"pin_budget_tokens": 1500,
|
|
750
|
+
"skills_index_budget_tokens": 800,
|
|
751
|
+
"recall_budget_tokens": 700
|
|
752
|
+
}
|
|
753
|
+
```
|
|
754
|
+
|
|
755
|
+
| Setting | Default | Description |
|
|
756
|
+
|---------|---------|-------------|
|
|
757
|
+
| `working_memory_budget_tokens` | 3000 | Total token ceiling for the whole `brain session-start` payload |
|
|
758
|
+
| `pin_budget_tokens` | 1500 | Sub-budget for pinned, always-present memories |
|
|
759
|
+
| `skills_index_budget_tokens` | 800 | Sub-budget for the advertised skills index (L0) |
|
|
760
|
+
| `recall_budget_tokens` | 700 | Sub-budget for context-relevant recalled memories |
|
|
761
|
+
|
|
762
|
+
Token counts use a dependency-free heuristic stored on each memory at write time. When pins exceed their budget they're selected by `--priority` then strength, and the overflow is reported rather than silently dropped.
|
|
493
763
|
|
|
494
|
-
|
|
764
|
+
## Benchmark
|
|
495
765
|
|
|
496
|
-
|
|
497
|
-
|----------|---------------|:---:|:---:|
|
|
498
|
-
| Multi-Session Continuity | Decisions carry from Session 1 → Session 2 | **+26.2%** | 0% |
|
|
499
|
-
| Cross-Agent Consistency | All agents follow memorized style guide | 0% | 0% |
|
|
500
|
-
| Accumulated Knowledge | 5 sessions of learning → better Session 6 | **+7.3%** | 0% |
|
|
501
|
-
| Error Pattern Learning | Past debugging → faster fix | **+4.8%** | 0% |
|
|
502
|
-
| Preference Retention | Preferences applied without re-stating | **+34.7%** | **+33.3%** |
|
|
766
|
+
Brain Memory ships with a controlled benchmark suite grounded in the 2025-2026 SOTA in long-term-memory evaluation. The system itself is a direct implementation of the [**CoALA**](https://arxiv.org/abs/2309.02427) agent-memory model (Sumers et al., 2023) — Pinned Tier maps to CoALA's *semantic* memory, procedural skills to *procedural* memory.
|
|
503
767
|
|
|
504
|
-
**
|
|
768
|
+
**Six scenarios**, each describable in one sentence:
|
|
505
769
|
|
|
506
|
-
|
|
|
507
|
-
|
|
508
|
-
|
|
|
509
|
-
|
|
|
510
|
-
|
|
|
770
|
+
| Id | Pitch | Tests |
|
|
771
|
+
|---|---|---|
|
|
772
|
+
| **A** *Noisy Project Folder* | "Your brain has 200 memories from 6 projects — does it find the 3 relevant ones?" | Retrieval under distractors (LongMemEval-S analog) |
|
|
773
|
+
| **B** *Three Sessions, One Decision* | "Postgres Monday, gRPC rewrite Wednesday, new resource Friday — still Postgres?" | Multi-session continuity + Pinned Tier ablation |
|
|
774
|
+
| **C** *The Contradiction Test* | "Tabs, then spaces, then tabs again — which version wins?" | Decay-weighted recency + contradiction handling |
|
|
775
|
+
| **D** *Skill Progressive Disclosure* | "Five skills indexed, one needed — does brain load just the one?" | CoALA Phase-2 L0/L1/L2 token efficiency |
|
|
776
|
+
| **E** *Continual Coding* | "Five bugs in order — does bug 5 finish faster than bug 1?" | Forward transfer; agent writes memories between tasks |
|
|
777
|
+
| **F** *Abstention* | "No deployment target in memory — does the agent ask or invent?" | Confabulation resistance |
|
|
511
778
|
|
|
512
|
-
**
|
|
779
|
+
**Methodology highlights** — agent under test is **DeepSeek V4 Pro** (single-shot), graded by a **cross-family judge panel** (Gemini + Gemma-4 + Qwen-3.5, majority vote — no judge shares the agent's family, [Preference Leakage 2502.01534](https://arxiv.org/abs/2502.01534), [PoLL 2404.18796](https://arxiv.org/abs/2404.18796)); deterministic distractor haystacks up to 1000 memories; real `brain session-start` / `brain recall` integration; tokens-per-successful-task as the headline efficiency metric; every scenario anchored by a no-memory floor and an oracle ceiling.
|
|
513
780
|
|
|
514
|
-
|
|
515
|
-
|-------|:---:|:---:|:---|
|
|
516
|
-
| Claude Code | 0.866 / PASS | 0.359 / **FAIL** | Brain prevented failure |
|
|
517
|
-
| Gemini CLI | 0.800 / PASS | 0.534 / PASS | +49.8% consistency |
|
|
518
|
-
| Codex CLI | 0.934 / PASS | 0.666 / PASS | +40.2% consistency |
|
|
781
|
+
> Results (DeepSeek V4 Pro, 3 runs/arm, June 2026). **Scenario A — retrieval under 1000 distractors:** `brain-full` is the only retriever whose memories let the model succeed — it passes **100%** while both **BM25 (0%)** and a **vector store (0%)** fail to surface the oracle memories — and it does so at the **leanest tokens-per-success (3,199)** of any passing arm. **Scenario D — skills:** loading only the relevant skill passes 100% at **1,580 tok/success, ~31% leaner** than dumping all skill bodies. Scenarios B (continuity) shows brain efficient-and-correct; C (contradiction) and F (abstention) are honest **nulls** — the base model handles them without memory. At n=3 the token gaps are directional, not yet significant; the pass-rate gradient is the result. Full tables: [brainmemory.ai/docs/benchmarks/results](https://brainmemory.ai/docs/benchmarks/results).
|
|
519
782
|
|
|
520
|
-
|
|
783
|
+
**Retrieval calibration probe (in-vivo, July 2026).** Separate from the controlled
|
|
784
|
+
suite: a 20-query retrieval probe on a real working brain (177 memories, 1,245
|
|
785
|
+
association edges; tag-derived topical queries with known targets) measured the
|
|
786
|
+
scoring-calibration fix shipped after beta.30. Before: hit@1 90%, MRR 0.950 — and a
|
|
787
|
+
four-term nonsense control query still returned ten results, the top scoring a
|
|
788
|
+
confident-looking 0.679. After query-coverage scaling and the relevance floor:
|
|
789
|
+
**hit@1 100% / hit@3 100% / MRR 1.000**, the nonsense control returns a single result
|
|
790
|
+
with honestly low relevance (0.213), and recall latency stays at ~46 ms (p50).
|
|
791
|
+
*Caveats: n=20, single brain, tag-derived queries favor the index's tag field, not
|
|
792
|
+
judge-graded — this is a calibration probe, not the controlled benchmark above.*
|
|
793
|
+
Method and detail: [brainmemory.ai/docs/benchmarks/results](https://brainmemory.ai/docs/benchmarks/results).
|
|
521
794
|
|
|
522
795
|
Run the benchmarks yourself:
|
|
523
796
|
|
|
@@ -528,24 +801,52 @@ npm test # Unit tests
|
|
|
528
801
|
node harness/runner.js # Full benchmark (cloud APIs)
|
|
529
802
|
```
|
|
530
803
|
|
|
531
|
-
|
|
804
|
+
Live methodology and per-scenario detail: [brainmemory.ai/docs/benchmarks](https://brainmemory.ai/docs/benchmarks) · Source: [`benchmark/`](benchmark/)
|
|
532
805
|
|
|
533
806
|
## Contributing
|
|
534
807
|
|
|
535
808
|
Contributions are welcome! See [CONTRIBUTING.md](CONTRIBUTING.md) for development setup, project structure, and how to submit changes.
|
|
536
809
|
|
|
537
|
-
##
|
|
538
|
-
|
|
539
|
-
|
|
540
|
-
|
|
541
|
-
|
|
542
|
-
|
|
543
|
-
- [
|
|
544
|
-
- [
|
|
545
|
-
- [
|
|
546
|
-
- [
|
|
547
|
-
- [
|
|
548
|
-
- [
|
|
810
|
+
## References
|
|
811
|
+
|
|
812
|
+
Brain Memory's architecture and benchmark methodology are grounded in the following work.
|
|
813
|
+
|
|
814
|
+
### Foundations — the agent-memory model Brain implements
|
|
815
|
+
|
|
816
|
+
- [**CoALA — Cognitive Architectures for Language Agents**](https://arxiv.org/abs/2309.02427) (arxiv 2309.02427) — Sumers, Yao, Narasimhan, Griffiths. The agent-memory taxonomy Brain implements directly. Pinned Tier → semantic memory, Skills → procedural memory, session-start aggregator → working memory.
|
|
817
|
+
- [**MemGPT — LLMs as Operating Systems**](https://arxiv.org/abs/2310.08560) (arxiv 2310.08560) — Packer et al. Paging-style memory management.
|
|
818
|
+
- [**Generative Agents — Interactive Simulacra of Human Behavior**](https://arxiv.org/abs/2304.03442) (arxiv 2304.03442) — Park et al. Recency · importance · relevance retrieval blend.
|
|
819
|
+
- [**Memory in the Age of AI Agents**](https://arxiv.org/abs/2512.13564) (arxiv 2512.13564) — Comprehensive survey on agent memory architectures.
|
|
820
|
+
- [**Mem0**](https://arxiv.org/abs/2504.19413) (arxiv 2504.19413) — Human-like memory reinforcement and decay.
|
|
821
|
+
- [**MemOS**](https://arxiv.org/abs/2507.03724) (arxiv 2507.03724) — Memory lifecycle state management.
|
|
822
|
+
|
|
823
|
+
### Memory benchmarks (the suite this work follows)
|
|
824
|
+
|
|
825
|
+
- [**LongMemEval**](https://arxiv.org/abs/2410.10813) (arxiv 2410.10813) — distractor haystacks (S / M / Oracle), abstention category, GPT-4o judge with 97% human agreement.
|
|
826
|
+
- [**MemoryAgentBench**](https://arxiv.org/abs/2507.05257) (arxiv 2507.05257) — four-competency framework; FactConsolidation inspired Scenario C.
|
|
827
|
+
- [**SWE-Bench-CL**](https://arxiv.org/abs/2507.00014) (arxiv 2507.00014) — repo-scoped chronological evaluation; template for Scenario E.
|
|
828
|
+
- [**LoCoMo**](https://arxiv.org/abs/2402.17753) (arxiv 2402.17753) — long-conversation memory benchmark.
|
|
829
|
+
- [**MIRIX**](https://arxiv.org/abs/2507.07957) (arxiv 2507.07957) — realistic synthetic-but-grounded memory benchmarks.
|
|
830
|
+
|
|
831
|
+
### Methodology — judging and benchmark hygiene
|
|
832
|
+
|
|
833
|
+
- [**Preference Leakage in LLM-as-judge**](https://arxiv.org/abs/2502.01534) (arxiv 2502.01534) — drives cross-family judging.
|
|
834
|
+
- [**When Judgment Becomes Noise — position bias**](https://arxiv.org/abs/2509.20293) (arxiv 2509.20293) — drives position-swap mitigation.
|
|
835
|
+
- [**Silent Judge — shortcut bias**](https://arxiv.org/abs/2509.26072) (arxiv 2509.26072) — drives rubric-based judging.
|
|
836
|
+
- [**LastingBench**](https://arxiv.org/abs/2506.21614) (arxiv 2506.21614) — benchmark-leakage defense.
|
|
837
|
+
|
|
838
|
+
### Neuroscience
|
|
839
|
+
|
|
840
|
+
- [**Ebbinghaus forgetting curve**](https://en.wikipedia.org/wiki/Forgetting_curve) — exponential memory decay.
|
|
841
|
+
- [**Spreading activation**](https://en.wikipedia.org/wiki/Spreading_activation) — Collins & Loftus network model of semantic memory.
|
|
842
|
+
- [**Hebbian theory**](https://en.wikipedia.org/wiki/Hebbian_theory) — "neurons that fire together wire together."
|
|
843
|
+
- [**Synaptic homeostasis hypothesis**](https://doi.org/10.1016/j.neuron.2013.10.024) — Tononi & Cirelli's theory of sleep function.
|
|
844
|
+
- [**SM-2 algorithm**](https://en.wikipedia.org/wiki/SuperMemo#Description_of_SM-2_algorithm) — spaced repetition scheduling.
|
|
845
|
+
- [**Memory reconsolidation**](https://en.wikipedia.org/wiki/Memory_consolidation#Reconsolidation) — recalling memories makes them temporarily malleable.
|
|
846
|
+
|
|
847
|
+
### Related projects
|
|
848
|
+
|
|
849
|
+
- [get-shit-done](https://github.com/gsd-build/get-shit-done) — Structured file-based workflow orchestration for AI agents.
|
|
549
850
|
|
|
550
851
|
## License
|
|
551
852
|
|