remem-mcp 0.7.0 → 0.7.3
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/README.md +79 -218
- package/dist/{chunk-OYUNKUZK.js → chunk-4WCDAEPP.js} +788 -11
- package/dist/chunk-4WCDAEPP.js.map +1 -0
- package/dist/{chunk-TGLOMFIE.js → chunk-ITMR6THY.js} +47 -7
- package/dist/chunk-ITMR6THY.js.map +1 -0
- package/dist/{chunk-ZXI67YIV.js → chunk-YU7WBJU5.js} +15 -3
- package/dist/{chunk-ZXI67YIV.js.map → chunk-YU7WBJU5.js.map} +1 -1
- package/dist/{engine-3JUSB3RK.js → engine-S47HHGAO.js} +2 -2
- package/dist/index.js +347 -39
- package/dist/index.js.map +1 -1
- package/dist/{resolver-YIFWRNMS.js → resolver-UHU4MBQX.js} +50 -15
- package/dist/resolver-UHU4MBQX.js.map +1 -0
- package/dist/sdk.js +3 -3
- package/dist/{server-SUKPVQWL.js → server-5PWQTRGZ.js} +2 -2
- package/dist/storage/schema.sql +63 -2
- package/package.json +1 -1
- package/scripts/verify-codegraph-stats.mjs +122 -0
- package/skills/remem-mcp/SKILL.md +11 -3
- package/dist/chunk-OYUNKUZK.js.map +0 -1
- package/dist/chunk-TGLOMFIE.js.map +0 -1
- package/dist/resolver-YIFWRNMS.js.map +0 -1
- /package/dist/{engine-3JUSB3RK.js.map → engine-S47HHGAO.js.map} +0 -0
- /package/dist/{server-SUKPVQWL.js.map → server-5PWQTRGZ.js.map} +0 -0
package/README.md
CHANGED
|
@@ -5,25 +5,11 @@
|
|
|
5
5
|
[](https://opensource.org/licenses/MIT)
|
|
6
6
|
[](https://github.com/tinhien11/remem-mcp)
|
|
7
7
|
|
|
8
|
-
> Your coding agent stops repeating the same mistakes
|
|
8
|
+
> Your coding agent stops repeating the same mistakes.
|
|
9
9
|
|
|
10
|
-
Local memory that survives context compaction
|
|
10
|
+
Local-first memory that survives context compaction. Learns from every error, injects fixes before the next attempt, and syncs to your git repo so your whole team shares it.
|
|
11
11
|
|
|
12
|
-
**
|
|
13
|
-
|
|
14
|
-
<video src="https://raw.githubusercontent.com/tinhien11/remem-mcp/main/docs/screenshots/demo-learning-loop.mp4" controls muted width="100%"></video>
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
|
|
18
|
-
### See it in action
|
|
19
|
-
|
|
20
|
-
| Error learning loop | CodeGraph search | Web viewer |
|
|
21
|
-
|---|---|---|
|
|
22
|
-
|  |  |  |
|
|
23
|
-
|
|
24
|
-
| Viewer: overview | CodeGraph: callers | CodeGraph: search |
|
|
25
|
-
|---|---|---|
|
|
26
|
-
|  |  |  |
|
|
12
|
+
**No API key. No cloud. No database server. Just a SQLite file.**
|
|
27
13
|
|
|
28
14
|
---
|
|
29
15
|
|
|
@@ -33,102 +19,87 @@ Local memory that survives context compaction — learns from every error, injec
|
|
|
33
19
|
npx remem-mcp setup
|
|
34
20
|
```
|
|
35
21
|
|
|
36
|
-
|
|
22
|
+
Auto-detects Claude Code, Cursor, Devin, Codex. Registers MCP server + hooks. Restart your agent.
|
|
23
|
+
|
|
24
|
+
That's it. Use your agent normally — memory works automatically.
|
|
37
25
|
|
|
38
26
|
```bash
|
|
39
|
-
npx remem-mcp
|
|
40
|
-
npx remem-mcp demo-codegraph # Live CodeGraph demo on facebook/react
|
|
41
|
-
npx remem-mcp status # One dashboard: everything at a glance
|
|
27
|
+
npx remem-mcp status # verify: hooks ✓, DB ✓, CodeGraph ✓
|
|
42
28
|
```
|
|
43
29
|
|
|
44
|
-
The demo creates a real TypeScript project, runs real `npm run build`, captures real TS2307 errors, and shows the full learning loop — capture → inject → fix → zero retries. No hardcoded strings.
|
|
45
|
-
|
|
46
30
|
---
|
|
47
31
|
|
|
48
|
-
##
|
|
32
|
+
## What happens automatically
|
|
49
33
|
|
|
50
|
-
|
|
34
|
+
| When | What |
|
|
35
|
+
|---|---|
|
|
36
|
+
| Session start | Past errors, decisions, and persona injected into agent context |
|
|
37
|
+
| Each prompt | Matching memory injected (you'll see `[remem-mcp]` at the top) |
|
|
38
|
+
| Tool calls | Verbose output offloaded to refs, Mermaid canvas injected (92% token cut) |
|
|
39
|
+
| Session end | Worker auto-extracts facts, consolidates summaries, updates persona |
|
|
51
40
|
|
|
52
|
-
|
|
53
|
-
- **Session start** → past errors, decisions, and persona injected into agent context
|
|
54
|
-
- **Each prompt** → matching memory injected (you'll see `[remem-mcp]` at the top)
|
|
55
|
-
- **Session end** → worker auto-extracts facts, consolidates summaries, updates persona
|
|
56
|
-
|
|
57
|
-
```bash
|
|
58
|
-
npx remem-mcp status # verify: hooks ✓, DB ✓, CodeGraph ✓
|
|
59
|
-
```
|
|
41
|
+
You don't run any commands. The agent calls `recall()` before answering and `capture()` after work — the skill tells it to.
|
|
60
42
|
|
|
61
43
|
---
|
|
62
44
|
|
|
63
|
-
##
|
|
45
|
+
## How it works
|
|
64
46
|
|
|
65
47
|
```
|
|
66
|
-
|
|
67
|
-
│
|
|
68
|
-
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
|
|
72
|
-
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
|
|
81
|
-
└────────┬─────────┘ └─────────┬──────────┘
|
|
82
|
-
│ │
|
|
83
|
-
▼ ▼
|
|
84
|
-
┌─────────────────────────────────────────────────┐
|
|
85
|
-
│ SQLite (memory.db) │
|
|
86
|
-
│ │
|
|
87
|
-
│ L0 captures ──▶ L1 atoms ──▶ L2 scenarios ──▶ L3 persona
|
|
88
|
-
│ (raw) (facts) (summaries) (preferences)
|
|
89
|
-
│ │
|
|
90
|
-
│ CodeGraph: symbols + calls + imports │
|
|
91
|
-
│ Wiki: markdown docs + ADRs │
|
|
92
|
-
│ Canvas: Mermaid nodes + edges + refs │
|
|
93
|
-
│ Skills: trigger conditions + steps + validation │
|
|
94
|
-
└───────────────────────────────────────────────────┘
|
|
48
|
+
AI Agent (Claude Code / Devin / Cursor / Codex)
|
|
49
|
+
│
|
|
50
|
+
├── MCP tools ──▶ recall, capture, codegraph_*, wiki_*, feedback
|
|
51
|
+
│
|
|
52
|
+
└── Hooks ──▶ SessionStart, UserPromptSubmit, PreToolUse,
|
|
53
|
+
PostToolUse, Stop, PostCompact
|
|
54
|
+
│
|
|
55
|
+
▼
|
|
56
|
+
SQLite (memory.db)
|
|
57
|
+
|
|
58
|
+
L0 captures → L1 atoms → L2 scenarios → L3 persona
|
|
59
|
+
(raw) (facts) (summaries) (preferences)
|
|
60
|
+
|
|
61
|
+
CodeGraph: symbols + calls + imports (tree-sitter, 9 languages)
|
|
62
|
+
Memory links: Hebbian co-retrieval (frequently co-retrieved = stronger)
|
|
95
63
|
```
|
|
96
64
|
|
|
97
|
-
|
|
65
|
+
**No LLM API key needed** — rule-based extraction + keyword grouping.
|
|
98
66
|
|
|
99
|
-
|
|
100
|
-
- **L1 atoms** — distilled facts with confidence scores (agent writes or worker extracts)
|
|
101
|
-
- **L2 scenarios** — auto-consolidated summaries when 5+ atoms share a topic
|
|
102
|
-
- **L3 persona** — user preferences auto-detected from repeated tags (2+ occurrences)
|
|
67
|
+
---
|
|
103
68
|
|
|
104
|
-
|
|
69
|
+
## CodeGraph
|
|
105
70
|
|
|
106
|
-
|
|
71
|
+
Structural code indexing via tree-sitter. The agent uses `codegraph_search` instead of grep to find symbols.
|
|
107
72
|
|
|
108
|
-
|
|
73
|
+
```bash
|
|
74
|
+
npx remem-mcp index --path src # index a directory
|
|
75
|
+
npx remem-mcp search-code --query "parseTar" # find symbols
|
|
76
|
+
npx remem-mcp callers <id> # who calls this?
|
|
77
|
+
npx remem-mcp impact <id> # blast radius
|
|
78
|
+
```
|
|
109
79
|
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
|
|
80
|
+
9 languages: TS/JS/Python/Go/Rust/Java/C/C++/C#. 6-strategy call resolution (import-map → same-module → unique-name → suffix → fuzzy). Stdlib calls filtered out.
|
|
81
|
+
|
|
82
|
+
| Repo | Files | Symbols | Calls | Time |
|
|
83
|
+
|---|---|---|---|---|
|
|
84
|
+
| remem-mcp | 79 | 301 | 6,456 | 3s |
|
|
85
|
+
| AZR Go | 455 | 3,417 | 41,603 | 111s |
|
|
86
|
+
| Orca TS | 3,000 | 7,632 | 78,981 | 705s |
|
|
114
87
|
|
|
115
88
|
---
|
|
116
89
|
|
|
117
90
|
## Why it's different
|
|
118
91
|
|
|
119
|
-
| | remem-mcp | Mem0 | Claude MEMORY.md | Mneme |
|
|
120
|
-
|
|
121
|
-
| **Survives compaction** | Yes
|
|
122
|
-
| **Learns from errors** | Yes — auto
|
|
123
|
-
| **
|
|
124
|
-
| **Memory
|
|
125
|
-
| **
|
|
126
|
-
| **
|
|
127
|
-
| **
|
|
128
|
-
| **
|
|
129
|
-
| **
|
|
130
|
-
| **API key** | No | Yes | No | No | No |
|
|
131
|
-
| **Cost** | Free | $19–249/mo | Free | Free | Free |
|
|
92
|
+
| | remem-mcp | Mem0 | Claude MEMORY.md | Mneme |
|
|
93
|
+
|---|---|---|---|---|
|
|
94
|
+
| **Survives compaction** | Yes | Yes — cloud | No — 200-line cap | Yes |
|
|
95
|
+
| **Learns from errors** | Yes — auto | No | No | No |
|
|
96
|
+
| **Search** | Hybrid BM25 + vector + entities | Vector only | No | Vector + graph |
|
|
97
|
+
| **Memory links** | Hebbian co-retrieval | No | No | Graph |
|
|
98
|
+
| **Decay/forget** | Yes | No | No | No |
|
|
99
|
+
| **CodeGraph** | Yes — 6-strategy call resolution | No | No | No |
|
|
100
|
+
| **Token offload** | Yes — Mermaid canvas | No | No | No |
|
|
101
|
+
| **Setup** | 1 command | API key + cloud | Built-in | Build from source |
|
|
102
|
+
| **Cost** | Free | $19–249/mo | Free | Free |
|
|
132
103
|
|
|
133
104
|
---
|
|
134
105
|
|
|
@@ -175,100 +146,20 @@ Add to `~/.codex/config.toml`:
|
|
|
175
146
|
[mcp_servers.remem-mcp]
|
|
176
147
|
command = "npx"
|
|
177
148
|
args = ["-y", "remem-mcp"]
|
|
178
|
-
|
|
179
|
-
[mcp_servers.remem-mcp.env]
|
|
180
|
-
REMEM_GLOBAL_SESSION_KEY = "global"
|
|
181
149
|
```
|
|
182
|
-
|
|
183
150
|
Then run `npx remem-mcp install-hooks`.
|
|
184
|
-
> MCP tools require `sandbox_mode = "danger-full-access"`.
|
|
185
151
|
</details>
|
|
186
152
|
|
|
187
153
|
---
|
|
188
154
|
|
|
189
|
-
##
|
|
190
|
-
|
|
191
|
-
Memory lives in a local SQLite database — outside the agent's context window. When the agent compacts or starts a new session, memory is re-injected automatically. No more re-explaining what you already told it yesterday.
|
|
192
|
-
|
|
193
|
-
**PreCompact hook**: when the agent is about to compact context, remem-mcp saves a checkpoint (decisions made, approaches tried, what's verified working) to the DB. After compaction, the agent recalls it — so the compact doesn't destroy your session's learnings.
|
|
194
|
-
|
|
195
|
-
Two layers: **automatic** (runs via hooks, zero tool calls) and **on-demand** (you call when you need deeper context).
|
|
196
|
-
|
|
197
|
-
### Automatic — three learning loops + compaction survival + symbolic memory
|
|
198
|
-
|
|
199
|
-
All run via lifecycle hooks. The agent doesn't need to call any tool.
|
|
200
|
-
|
|
201
|
-
1. **Error learning** — command fails → capture → inject fix before next attempt → succeed → upvote.
|
|
202
|
-
|
|
203
|
-
2. **Decision learning** — `npm install`, `git commit`, config → auto-capture → inject past decisions before similar commands.
|
|
204
|
-
|
|
205
|
-
3. **Pattern learning** — Write/Edit → auto-capture code patterns → inject same-language patterns before editing.
|
|
206
|
-
|
|
207
|
-
4. **Compaction survival** — PreCompact hook fires before context compaction → saves checkpoint → agent recalls after compact. Memory survives.
|
|
208
|
-
|
|
209
|
-
5. **Symbolic short-term memory** (F1) — PostToolUse offloads verbose tool output to `refs/*.md` files and appends a node to a Mermaid canvas. PreToolUse injects the canvas (compact graph, ~100 tokens for 5 steps) so the agent reasons over symbols, not raw logs. Drill down via `ref_read(node_id)`. **92% token reduction** vs. raw tool logs. Enable with `REMEM_FLOW=full`.
|
|
210
|
-
|
|
211
|
-
6. **Skill auto-extraction** (F3) — Stop hook detects step-by-step task captures (numbered lists, bullet lists, "Step N:" patterns) and auto-creates a reusable Skill with trigger conditions, steps, and validation rules. Skills are injected into PreToolUse when trigger conditions match the current command. Archived skills are always injected. Enable with `REMEM_FLOW=full`.
|
|
212
|
-
|
|
213
|
-
### Memory Proxy — zero-code LLM integration (F2)
|
|
214
|
-
|
|
215
|
-
For agents that don't support MCP hooks (or any OpenAI/Anthropic client):
|
|
216
|
-
|
|
217
|
-
```bash
|
|
218
|
-
remem-mcp proxy # Starts HTTP proxy on :8765
|
|
219
|
-
```
|
|
220
|
-
|
|
221
|
-
Point your agent's base URL to `http://localhost:8765`. The proxy intercepts `/v1/chat/completions` (OpenAI) and `/v1/messages` (Anthropic), injects a `<remem-mcp>` memory block into the system prompt (recall + skills + canvas), forwards to the upstream LLM, and auto-captures the conversation. Session binding via `x-remem-team` / `x-remem-agent` headers or `POST /session/init`.
|
|
222
|
-
|
|
223
|
-
### On-demand — CodeGraph, Wiki, Search
|
|
224
|
-
|
|
225
|
-
When the automatic loops aren't enough, use these for deeper code navigation.
|
|
226
|
-
|
|
227
|
-
**CodeGraph** — symbol search, callers/callees, impact analysis. **Auto-indexes on first use** — just call `codegraph_search` and it indexes `src/` automatically. No manual `codegraph_index` needed.
|
|
228
|
-
|
|
229
|
-
```bash
|
|
230
|
-
# Search symbols (auto-indexes src/ on first call)
|
|
231
|
-
npx remem-mcp search-code --query "parseTar"
|
|
232
|
-
# → parseTar at src/parse.ts:22
|
|
233
|
-
|
|
234
|
-
# List symbols in a file
|
|
235
|
-
npx remem-mcp list-code src/reporters/fancy.ts
|
|
236
|
-
# → Class L49-135 FancyReporter
|
|
237
|
-
# → Method L86-134 formatLogObj
|
|
238
|
-
|
|
239
|
-
# Trace callers / callees / impact (use symbol ID from search)
|
|
240
|
-
npx remem-mcp callers 01KZXPPHF93TS4HV8FWCSSK36A
|
|
241
|
-
npx remem-mcp impact 01KZXPPHF93TS4HV8FWCSSK36A
|
|
242
|
-
|
|
243
|
-
# Manual re-index (only needed after major changes)
|
|
244
|
-
npx remem-mcp index --path src --repo .
|
|
245
|
-
|
|
246
|
-
# Wiki + viewer
|
|
247
|
-
npx remem-mcp wiki ingest --path docs # Index markdown docs + ADRs
|
|
248
|
-
npx remem-mcp wiki outdated # Find outdated wiki pages
|
|
249
|
-
npx remem-mcp viewer # Web UI at localhost:7331
|
|
250
|
-
```
|
|
251
|
-
|
|
252
|
-
- **CodeGraph** — symbol search, callers/callees, impact analysis. Auto-indexes on first `codegraph_search` call. Auto-scoped to your project.
|
|
253
|
-
- **Wiki** — markdown docs, ADRs, outdated detection.
|
|
254
|
-
- **Search** — hybrid BM25 + sqlite-vec vector search with RRF fusion. `explain_recall` shows scores.
|
|
255
|
-
|
|
256
|
-
|
|
257
|
-
|
|
258
|
-
---
|
|
259
|
-
|
|
260
|
-
## Daily commands
|
|
155
|
+
## Useful commands
|
|
261
156
|
|
|
262
157
|
```bash
|
|
263
|
-
npx remem-mcp status #
|
|
264
|
-
npx remem-mcp viewer #
|
|
265
|
-
npx remem-mcp errors #
|
|
266
|
-
npx remem-mcp
|
|
267
|
-
npx remem-mcp
|
|
268
|
-
npx remem-mcp recent [N] # Recent captures
|
|
269
|
-
npx remem-mcp proxy # Start Memory Proxy (HTTP, :8765)
|
|
270
|
-
npx remem-mcp skill-extract # Batch extract skills from task captures
|
|
271
|
-
npx remem-mcp help all # Full list of 40+ subcommands
|
|
158
|
+
npx remem-mcp status # health + hooks + DB + CodeGraph
|
|
159
|
+
npx remem-mcp viewer # web UI at localhost:7331
|
|
160
|
+
npx remem-mcp errors # error dashboard
|
|
161
|
+
npx remem-mcp recent [N] # recent captures
|
|
162
|
+
npx remem-mcp help all # full list of 40+ subcommands
|
|
272
163
|
```
|
|
273
164
|
|
|
274
165
|
---
|
|
@@ -280,60 +171,30 @@ All settings have defaults. Config file is optional: `~/.config/remem-mcp/config
|
|
|
280
171
|
| Setting | Env var | Default |
|
|
281
172
|
|---|---|---|
|
|
282
173
|
| DB path | `REMEM_DB_PATH` | `~/.local/share/remem-mcp/memory.db` |
|
|
283
|
-
| **Unified flow** (F1+F3) | `REMEM_FLOW` | _(unset, set to `full`)_ |
|
|
284
|
-
| **Symbolic memory** (F1) | `REMEM_OFFLOAD_ENABLED` | `false` (set to `true`) |
|
|
285
|
-
| **Pipeline** (F1/F3) | `REMEM_PIPELINE` | `noop` (`mermaid`, `skill`, `llm-mermaid`) |
|
|
286
|
-
| **Proxy port** (F2) | `REMEM_PROXY_PORT` | `8765` |
|
|
287
|
-
| **Proxy upstream** (F2) | `REMEM_UPSTREAM_URL` | `https://api.openai.com` |
|
|
288
|
-
| **Proxy API key** (F2) | `REMEM_UPSTREAM_API_KEY` | _(from OPENAI_API_KEY)_ |
|
|
289
174
|
| Cross-project memory | `REMEM_GLOBAL_SESSION_KEY` | _(unset)_ |
|
|
290
|
-
|
|
|
291
|
-
| Auto-global classification | `auto_global=true` on capture | _(off)_ |
|
|
175
|
+
| Unified flow (F1+F2+F3) | `REMEM_FLOW` | _(unset, set to `full`)_ |
|
|
292
176
|
| Suppress hook feedback | `REMEM_QUIET` | _(unset, set to `1`)_ |
|
|
293
|
-
| Retro window (days) | `REMEM_RETRO_DAYS` | `7` |
|
|
294
|
-
| Core-only mode (disable advanced tools) | `REMEM_CORE_ONLY` | _(unset, set to `1`)_ |
|
|
295
|
-
| LLM API key (pipeline) | `REMEM_LLM_API_KEY` | _(unset)_ |
|
|
296
|
-
|
|
297
|
-
**Team sharing** — `npx remem-mcp sync-export` writes `.remem-mcp/memory-export.jsonl`. Commit it to git. Team members get the same memory on `git pull` (auto-imports on startup).
|
|
298
|
-
|
|
299
|
-
---
|
|
300
177
|
|
|
301
|
-
|
|
178
|
+
**Team sharing** — `npx remem-mcp sync-export` writes `.remem-mcp/memory-export.jsonl`. Commit it to git. Team members get the same memory on `git pull`.
|
|
302
179
|
|
|
303
|
-
|
|
304
|
-
|
|
305
|
-
|
|
306
|
-
|
|
307
|
-
await memory.capture("We chose SQLite for storage.", "decision", ["arch"]);
|
|
308
|
-
const results = await memory.recall("storage decision");
|
|
180
|
+
**Per-repo capture exclusions** — Drop a `.remem.toml` in any project root:
|
|
181
|
+
```toml
|
|
182
|
+
[capture]
|
|
183
|
+
ignore_paths = ["node_modules", "dist", ".git", "*.min.js"]
|
|
309
184
|
```
|
|
310
185
|
|
|
311
186
|
---
|
|
312
187
|
|
|
313
188
|
## Benchmark
|
|
314
189
|
|
|
315
|
-
remem-mcp
|
|
316
|
-
|
|
317
|
-
|
|
318
|
-
|
|
319
|
-
|
|
|
320
|
-
|
|
321
|
-
| **AMB Layer 1** (basic recall) | **100** | — | — | — |
|
|
322
|
-
| **AMB Layer 2** (multi-session) | **100** | — | — | — |
|
|
323
|
-
| **AMB Layer 3** (scale + distractors) | **100** | — | — | — |
|
|
324
|
-
| **LoCoMo** (long conversation QA) | **95** | — | 92.5 | — |
|
|
325
|
-
| **PersonaMem** (personalization) | **100** | 76 | — | 48 |
|
|
326
|
-
| **LongMemEval** (long-term memory, ICR 2025) | **96** | — | 94.4 | — |
|
|
327
|
-
|
|
328
|
-
- **PersonaMem** — [bowen-upenn/PersonaMem](https://github.com/bowen-upenn/PersonaMem) (588 questions, 20 personas, multiple-choice QA). TencentDB reports 76% with memory enabled, 48% without. remem-mcp scores **100%** using a search-recall proxy (no LLM API key needed).
|
|
329
|
-
- **LoCoMo** — long conversation multi-hop QA (19 sessions, 400+ turns). Mem0 reports 92.5%. remem-mcp scores **95%** with keyword + semantic-similarity scoring.
|
|
330
|
-
- **AMB** — Agent Memory Benchmark (L1: 56 recall tests, L2: 5 multi-session scenarios, L3: 1K+ memories with distractors). remem-mcp scores **100/100/100** using the official AMB CLI.
|
|
331
|
-
- **LongMemEval** — [xiaowu0162/LongMemEval](https://github.com/xiaowu0162/LongMemEval) (ICLR 2025, 500 questions, 5 memory abilities: temporal reasoning, multi-session, knowledge update, single-session recall, abstention). Mem0 reports 94.4%. remem-mcp scores **96%** on the oracle variant.
|
|
332
|
-
|
|
333
|
-
Run the benchmarks:
|
|
190
|
+
| Benchmark | remem-mcp | Mem0 | Without memory |
|
|
191
|
+
|---|---|---|---|
|
|
192
|
+
| **AMB** (L1/L2/L3) | **100/100/100** | — | — |
|
|
193
|
+
| **LoCoMo** (long conversation QA) | **95** | 92.5 | — |
|
|
194
|
+
| **PersonaMem** (personalization) | **100** | — | 48 |
|
|
195
|
+
| **LongMemEval** (ICLR 2025) | **96** | 94.4 | — |
|
|
334
196
|
|
|
335
197
|
```bash
|
|
336
|
-
bash scripts/bench-all.sh # Full: AMB + LoCoMo + PersonaMem (~5 min)
|
|
337
198
|
bash scripts/bench-all.sh --quick # AMB only (~2 min)
|
|
338
199
|
```
|
|
339
200
|
|
|
@@ -341,11 +202,11 @@ bash scripts/bench-all.sh --quick # AMB only (~2 min)
|
|
|
341
202
|
|
|
342
203
|
## Architecture
|
|
343
204
|
|
|
344
|
-
See [ARCHITECTURE.md](./ARCHITECTURE.md) for full system diagrams, schema, and performance
|
|
205
|
+
See [ARCHITECTURE.md](./ARCHITECTURE.md) for full system diagrams, schema, and performance details.
|
|
345
206
|
|
|
346
207
|
## Credits
|
|
347
208
|
|
|
348
|
-
Core based on [TencentDB Agent Memory](https://github.com/TencentCloud/TencentDB-Agent-Memory) (MIT, Tencent 2026).
|
|
209
|
+
Core based on [TencentDB Agent Memory](https://github.com/TencentCloud/TencentDB-Agent-Memory) (MIT, Tencent 2026). CodeGraph call resolution adapted from Codebase-Memory (arXiv:2603.27277). Recall boost adapted from [ai-memory](https://github.com/akitaonrails/ai-memory) by Akita On Rails. Contextual retrieval from [Anthropic](https://www.anthropic.com/news/contextual-retrieval) (2024).
|
|
349
210
|
|
|
350
211
|
## License
|
|
351
212
|
|