apiary-memory 0.2.0__tar.gz
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.
- apiary_memory-0.2.0/LICENSE +21 -0
- apiary_memory-0.2.0/PKG-INFO +333 -0
- apiary_memory-0.2.0/README.md +309 -0
- apiary_memory-0.2.0/pyproject.toml +53 -0
- apiary_memory-0.2.0/setup.cfg +4 -0
- apiary_memory-0.2.0/src/apiary_memory.egg-info/PKG-INFO +333 -0
- apiary_memory-0.2.0/src/apiary_memory.egg-info/SOURCES.txt +14 -0
- apiary_memory-0.2.0/src/apiary_memory.egg-info/dependency_links.txt +1 -0
- apiary_memory-0.2.0/src/apiary_memory.egg-info/entry_points.txt +5 -0
- apiary_memory-0.2.0/src/apiary_memory.egg-info/top_level.txt +1 -0
- apiary_memory-0.2.0/src/hive_mcp/__init__.py +27 -0
- apiary_memory-0.2.0/src/hive_mcp/cli.py +261 -0
- apiary_memory-0.2.0/src/hive_mcp/stdio.py +799 -0
- apiary_memory-0.2.0/src/hive_mcp/store.py +872 -0
- apiary_memory-0.2.0/src/hive_mcp/tools.json +211 -0
- apiary_memory-0.2.0/tests/test_bridge.py +1437 -0
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 Apiary
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
|
@@ -0,0 +1,333 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: apiary-memory
|
|
3
|
+
Version: 0.2.0
|
|
4
|
+
Summary: Local-first memory your AI agents share over MCP. Ships a working SQLite memory store; agents propose, only you confirm.
|
|
5
|
+
Author: Apiary
|
|
6
|
+
License-Expression: MIT
|
|
7
|
+
Project-URL: Homepage, https://hello-fengchao.com
|
|
8
|
+
Project-URL: Repository, https://github.com/330899935/apiary-memory
|
|
9
|
+
Project-URL: App (macOS), https://apps.apple.com/app/id6806992983
|
|
10
|
+
Project-URL: Source, https://github.com/330899935/apiary-memory
|
|
11
|
+
Project-URL: Changelog, https://github.com/330899935/apiary-memory/blob/main/CHANGELOG.md
|
|
12
|
+
Keywords: mcp,mcp-server,model-context-protocol,memory,ai-memory,local-first,privacy,ai-agents,claude,cursor
|
|
13
|
+
Classifier: Development Status :: 3 - Alpha
|
|
14
|
+
Classifier: Intended Audience :: Developers
|
|
15
|
+
Classifier: Operating System :: MacOS
|
|
16
|
+
Classifier: Operating System :: OS Independent
|
|
17
|
+
Classifier: Programming Language :: Python :: 3
|
|
18
|
+
Classifier: Topic :: Database
|
|
19
|
+
Classifier: Topic :: Software Development :: Libraries :: Python Modules
|
|
20
|
+
Requires-Python: >=3.8
|
|
21
|
+
Description-Content-Type: text/markdown
|
|
22
|
+
License-File: LICENSE
|
|
23
|
+
Dynamic: license-file
|
|
24
|
+
|
|
25
|
+
# Apiary Memory MCP
|
|
26
|
+
|
|
27
|
+
<!-- mcp-name: io.github.330899935/apiary-memory -->
|
|
28
|
+
|
|
29
|
+
> **Also known as:** Apiary Memory · 记忆蜂巢 · Memory Hive · 蜂巢 ·
|
|
30
|
+
> `io.github.330899935/apiary-memory` (MCP Registry name, from v0.2.0) ·
|
|
31
|
+
> `hive-memory` (Registry name up to v0.1.9) · `memory-hive-mcp` (PyPI package until the rename ships).
|
|
32
|
+
> Not affiliated with [hive.com](https://hive.com) (team collaboration) or the `hive-mcp` Claude Code plugin.
|
|
33
|
+
|
|
34
|
+
**Local-first personal memory for AI agents, over MCP.** Your agents can search what you have
|
|
35
|
+
confirmed and propose what they think you should keep; **only you can confirm it.** Everything
|
|
36
|
+
lives in one file on your own disk, and nothing is sent anywhere.
|
|
37
|
+
|
|
38
|
+
It runs on its own. A working SQLite memory store ships inside this package — no account, no
|
|
39
|
+
cloud, no daemon, no dependencies.
|
|
40
|
+
|
|
41
|
+
```bash
|
|
42
|
+
pip install memory-hive-mcp
|
|
43
|
+
```
|
|
44
|
+
|
|
45
|
+
**Desktop app, no terminal:** download [`memory-hive-mcp-0.1.9.mcpb`](https://github.com/330899935/apiary-memory/releases/download/v0.1.9/memory-hive-mcp-0.1.9.mcpb) and double-click it.
|
|
46
|
+
|
|
47
|
+
Measured on 2026-10-27, on one MacBook with the Hive running locally — script and raw output in [`bench/`](bench/):
|
|
48
|
+
|
|
49
|
+
| Measure | Result |
|
|
50
|
+
| --- | --- |
|
|
51
|
+
| Search latency | **p95 111 ms**, p50 76 ms (warm, includes HTTP) |
|
|
52
|
+
| Retrieval | correct entry in **top-5 for 97.5%** of 40 trials |
|
|
53
|
+
| Corpus | **1,303 entries · 2.39 M chars · 45 days**, plaintext only |
|
|
54
|
+
|
|
55
|
+
*Retrieval used each entry's own title as the query — an easy set. It shows the index is complete and ranks correctly on exact terms, not how well it answers a question asked in your own words. Run [`bench/bench_search.py`](bench/bench_search.py) on your own Hive.*
|
|
56
|
+
|
|
57
|
+
---
|
|
58
|
+
|
|
59
|
+
## What you get without the app
|
|
60
|
+
|
|
61
|
+
A complete, if modest, memory server. On a machine with nothing else installed:
|
|
62
|
+
|
|
63
|
+
| Call | What happens |
|
|
64
|
+
| --- | --- |
|
|
65
|
+
| `initialize` | answered locally, protocol `2024-11-05` |
|
|
66
|
+
| `tools/list` | answered locally, six tools |
|
|
67
|
+
| `propose` | **writes to a real database** on your disk |
|
|
68
|
+
| `recall` | **searches that database** and returns ranked entries |
|
|
69
|
+
| `get_context` | a digest of what you confirmed most recently |
|
|
70
|
+
| `manage` | read one entry, list what is pending, forget, restore |
|
|
71
|
+
| `say` / `board` | a shared message board for coordinating several agents |
|
|
72
|
+
|
|
73
|
+
Storage is SQLite at `~/.memory-hive/hive.db` by default. Search is lexical — character bigrams
|
|
74
|
+
for Chinese, words for Latin script — so it ranks well when the words overlap and finds nothing
|
|
75
|
+
when they do not. There are no embeddings and no vector index. Every response says
|
|
76
|
+
`"mode": "local"` so you can always tell which engine answered.
|
|
77
|
+
|
|
78
|
+
**Without the app, the database is plain SQLite: it is not encrypted.** That is stated here
|
|
79
|
+
rather than discovered later — see [Honest limits](#honest-limits).
|
|
80
|
+
|
|
81
|
+
## What the app adds
|
|
82
|
+
|
|
83
|
+
[Apiary Memory for macOS](https://apps.apple.com/app/id6806992983) is the same design with the
|
|
84
|
+
expensive parts built: eight vaults with four of them locked behind a PIN (Argon2id-wrapped key,
|
|
85
|
+
AES-256 at rest), a daily evolution log, multi-agent identity management, and a UI for reading
|
|
86
|
+
and confirming. It also answers MCP itself, with richer ranking.
|
|
87
|
+
|
|
88
|
+
If the app is running, this package uses it and leaves the local file alone. If it is not, the
|
|
89
|
+
local store answers. You do not have to choose up front:
|
|
90
|
+
|
|
91
|
+
| `HIVE_MODE` | Behaviour |
|
|
92
|
+
| --- | --- |
|
|
93
|
+
| `auto` (default) | Ask the app. Use the local store **only if the app is unreachable.** |
|
|
94
|
+
| `local` | Never touch the network. Always use the local store. |
|
|
95
|
+
| `remote` | Only the app. A connection failure is an error, not a reroute. |
|
|
96
|
+
|
|
97
|
+
One rule holds in every mode: **if the app is running and answers with an error, that error is
|
|
98
|
+
returned unchanged.** Falling back on a refusal would write your memory into a second database
|
|
99
|
+
while your client believed it had reached the app. Silent data forking is worse than an error
|
|
100
|
+
message, so a reply from the app is final. The fallback fires on a connection failure, and on
|
|
101
|
+
nothing else.
|
|
102
|
+
|
|
103
|
+
## Install
|
|
104
|
+
|
|
105
|
+
```bash
|
|
106
|
+
pip install memory-hive-mcp # the server: memory-hive-mcp
|
|
107
|
+
# the human's CLI: memory-hive-mcp-cli
|
|
108
|
+
```
|
|
109
|
+
|
|
110
|
+
Python 3.8 or newer. There is nothing else to install — the package uses only the standard
|
|
111
|
+
library, and `sqlite3` is part of it.
|
|
112
|
+
|
|
113
|
+
From a checkout:
|
|
114
|
+
|
|
115
|
+
```bash
|
|
116
|
+
git clone https://github.com/330899935/apiary-memory
|
|
117
|
+
cd apiary-memory
|
|
118
|
+
PYTHONPATH=src python -m hive_mcp.stdio
|
|
119
|
+
```
|
|
120
|
+
|
|
121
|
+
## Connect a client
|
|
122
|
+
|
|
123
|
+
Any MCP host that can run a stdio command will work. The minimum is the command; everything else
|
|
124
|
+
is optional.
|
|
125
|
+
|
|
126
|
+
**Claude Desktop** — `~/Library/Application Support/Claude/claude_desktop_config.json`:
|
|
127
|
+
|
|
128
|
+
```json
|
|
129
|
+
{
|
|
130
|
+
"mcpServers": {
|
|
131
|
+
"memory-hive": {
|
|
132
|
+
"command": "memory-hive-mcp",
|
|
133
|
+
"env": { "HIVE_MODE": "local" }
|
|
134
|
+
}
|
|
135
|
+
}
|
|
136
|
+
}
|
|
137
|
+
```
|
|
138
|
+
|
|
139
|
+
**Cursor** — `~/.cursor/mcp.json`, same `mcpServers` shape.
|
|
140
|
+
|
|
141
|
+
**Any host that only speaks HTTP** — start the app and point it at `http://127.0.0.1:8768/mcp`
|
|
142
|
+
with the headers `X-Hive-Agent` and `X-Hive-Token`. This package is the stdio path; the HTTP
|
|
143
|
+
endpoint belongs to the app.
|
|
144
|
+
|
|
145
|
+
## Prerequisite
|
|
146
|
+
|
|
147
|
+
**None, for the local store** — install the package and it works. A running app is a prerequisite
|
|
148
|
+
only for the app's own features: encrypted vaults, the evolution log, and multi-agent identity.
|
|
149
|
+
If you want those, install [the Hive app](https://apps.apple.com/app/id6806992983) (macOS,
|
|
150
|
+
version **1.4.5 or newer**) and leave `HIVE_MODE` at `auto`.
|
|
151
|
+
|
|
152
|
+
## The six tools
|
|
153
|
+
|
|
154
|
+
The six tools are the whole surface. They are the same six whether the app or the local store
|
|
155
|
+
answers; only the ranking behind them differs — and, for two inputs, whether they exist at all.
|
|
156
|
+
The app accepts precomputed embeddings (`recall.query_vector`, `propose.vector`); the local
|
|
157
|
+
store has no embedder, so `tools/list` omits those two wherever the local store can answer,
|
|
158
|
+
including under `auto` where the app might stop answering mid-session. A field that would be
|
|
159
|
+
accepted and silently ignored is not advertised at all.
|
|
160
|
+
|
|
161
|
+
The same rule covers wording. `recall.query` is described as the lexical ranker it is wherever
|
|
162
|
+
the local store can answer — words and character bigrams, no embeddings — and as the app's
|
|
163
|
+
index-based ranking with `query_vector` for meaning wherever only the app can answer. A
|
|
164
|
+
description is a claim about what will happen to the input, so it is narrowed with the inputs.
|
|
165
|
+
|
|
166
|
+
Defaults are not narrowed, they are shared. Omit `limit` and you get 10 from `recall` and 8 from
|
|
167
|
+
`get_context`; omit `compact` and bodies come back trimmed; omit `ttl` on `say` and the card
|
|
168
|
+
lives 24 hours. Both engines answer those numbers identically, because a default is a number in
|
|
169
|
+
the schema rather than a capability one engine lacks. `store.py` holds each one as a named
|
|
170
|
+
constant, and the tests compare the answer against the advertised number instead of a second
|
|
171
|
+
hand-written copy of it.
|
|
172
|
+
|
|
173
|
+
| Tool | Kind | What it is for |
|
|
174
|
+
| --- | --- | --- |
|
|
175
|
+
| `recall` | read | Search confirmed memory by topic. `{"query": "..."}` plus optional `limit`, `vaults`, `compact`. |
|
|
176
|
+
| `propose` | write | Submit a memory for the human to approve. Lands in a queue; changes no confirmed memory. |
|
|
177
|
+
| `manage` | read/write | Address one entry: `list_pending`, `get_by_id`, `forget`, `restore`, `discard_pending`. |
|
|
178
|
+
| `get_context` | read | A short digest of recent confirmed memory, to seed a conversation. |
|
|
179
|
+
| `say` | write | Post to the shared board so other agents see what was done or handed over. |
|
|
180
|
+
| `board` | read | Read the board, optionally only what arrived since your last cursor. |
|
|
181
|
+
|
|
182
|
+
Two properties are worth stating plainly, because they are the product:
|
|
183
|
+
|
|
184
|
+
- **`propose` succeeding means "queued", not "stored".** `recall` will not return it until a
|
|
185
|
+
human confirms it.
|
|
186
|
+
- **`manage` has no `approve` action.** There is no path from an agent to a confirmed memory.
|
|
187
|
+
That is not an omission; it is the design.
|
|
188
|
+
|
|
189
|
+
## Response shapes
|
|
190
|
+
|
|
191
|
+
`tools/list` is one surface for both engines, and an answer should be too: a model that has
|
|
192
|
+
to branch on which engine replied will eventually branch wrong. `board` therefore answers
|
|
193
|
+
the app's shape, field for field — `items`, `unread`, `count`, `cursor`, `reader` — and
|
|
194
|
+
`cursor` is the id of the newest card, the same kind of value the app returns.
|
|
195
|
+
|
|
196
|
+
The cursor is what a poll carries *across* an engine switch, which is why it gets the space.
|
|
197
|
+
Under `auto` one call can be answered by the app and the next by the local store. Before 0.1.9
|
|
198
|
+
the local engine answered a float epoch timestamp and read the app's card id with `float()`,
|
|
199
|
+
so a cursor fetched from the app came back as a full re-read of the board and nothing said
|
|
200
|
+
so; a float sent the other way matched no card id, and the app answered with its newest
|
|
201
|
+
`limit`, equally quietly.
|
|
202
|
+
|
|
203
|
+
A cursor can still fail to resolve: its card expired (cards live 24 hours by default), or it
|
|
204
|
+
was issued by the other engine, whose ids are `CB…` in the app and `msg_…` here. Neither
|
|
205
|
+
engine can resume from that, so both return the newest window — but this one also sets
|
|
206
|
+
`since_unknown` and a `note` saying which of those happened. `unread` counts every live card
|
|
207
|
+
after the cursor and `count` counts what the answer carries, so `unread > count` means the
|
|
208
|
+
page did not cover everything waiting, and the `note` says so in words.
|
|
209
|
+
|
|
210
|
+
Three shapes are **not** unified, because the engines really do hold different data:
|
|
211
|
+
|
|
212
|
+
- a `recall` entry names its vault `vault` from the local store and `wh` from the app;
|
|
213
|
+
- `get_context` returns one flat `digest` locally and one bucket per warehouse (`context`,
|
|
214
|
+
keyed by warehouse name) from the app;
|
|
215
|
+
- `roster` entries share `agent_id`; locally they also carry `messages` and `last_seen`,
|
|
216
|
+
while the app adds its registry card — display name, seat, leases, current task — which a
|
|
217
|
+
SQLite file has no way to know.
|
|
218
|
+
|
|
219
|
+
## Confirming what agents propose
|
|
220
|
+
|
|
221
|
+
Agents propose. You confirm. The confirmation lives in a terminal, in a command you have to
|
|
222
|
+
type — deliberately not a tool an agent can call.
|
|
223
|
+
|
|
224
|
+
```bash
|
|
225
|
+
memory-hive-mcp-cli list # what is waiting
|
|
226
|
+
memory-hive-mcp-cli show <id> # read one in full
|
|
227
|
+
memory-hive-mcp-cli approve <id> # confirm it, or --all
|
|
228
|
+
memory-hive-mcp-cli reject <id> # drop it for good
|
|
229
|
+
memory-hive-mcp-cli search "deploy key" # search confirmed memory
|
|
230
|
+
memory-hive-mcp-cli forget <id> # soft delete, reversible
|
|
231
|
+
memory-hive-mcp-cli restore <id> # undo forget
|
|
232
|
+
memory-hive-mcp-cli stats # counts and the database path
|
|
233
|
+
memory-hive-mcp-cli export --out dump.json
|
|
234
|
+
memory-hive-mcp-cli path
|
|
235
|
+
```
|
|
236
|
+
|
|
237
|
+
`--db <path>` on any command points at a different database. `HIVE_DB` does the same from the
|
|
238
|
+
environment.
|
|
239
|
+
|
|
240
|
+
## Where your memory lives
|
|
241
|
+
|
|
242
|
+
`~/.memory-hive/hive.db`, unless `HIVE_DB` says otherwise. One SQLite file; copy it to back it
|
|
243
|
+
up, delete it to start over.
|
|
244
|
+
|
|
245
|
+
The board keeps its messages in the same file, so there is only ever one thing to back up. Cards
|
|
246
|
+
expire: 24 hours unless `ttl` says otherwise, clamped to between 60 seconds and 30 days, which
|
|
247
|
+
is what the app does too — a board that grows forever stops being a board.
|
|
248
|
+
|
|
249
|
+
## How search works, and what it does not do
|
|
250
|
+
|
|
251
|
+
Chinese text is indexed as single characters plus overlapping character bigrams; Latin text as
|
|
252
|
+
words. A query is scored by how much of it the entry covers, then nudged by recency — a 10%
|
|
253
|
+
multiplier, enough to break ties and not enough to overturn relevance.
|
|
254
|
+
|
|
255
|
+
This is a ranker, not an understander.
|
|
256
|
+
|
|
257
|
+
- **Scores are relative.** `0.8` means "closer than `0.4` for this query", not "80% confident".
|
|
258
|
+
They are not comparable across queries.
|
|
259
|
+
- **No semantic recall.** Ask about "money" and an entry that only says "invoice" will not be
|
|
260
|
+
found. Related-word coverage needs an embedder; this package has no dependencies and does not
|
|
261
|
+
pretend otherwise.
|
|
262
|
+
- **The ranking is the one the description states.** `tools/list` says "words and character
|
|
263
|
+
bigrams ... no embeddings" here, and that sentence is true of this engine. It is narrowed per
|
|
264
|
+
engine like the embedding inputs are, because the app's answer to the same question differs.
|
|
265
|
+
- **A miss is reported as a miss.** When nothing matches, `recall` says so and mentions how many
|
|
266
|
+
entries are awaiting confirmation, rather than returning a plausible neighbour.
|
|
267
|
+
|
|
268
|
+
## Configuration
|
|
269
|
+
|
|
270
|
+
| Variable | Default | Purpose |
|
|
271
|
+
| --- | --- | --- |
|
|
272
|
+
| `HIVE_MODE` | `auto` | `auto`, `local`, or `remote` — see the table above |
|
|
273
|
+
| `HIVE_DB` | `~/.memory-hive/hive.db` | Where the local database lives |
|
|
274
|
+
| `HIVE_MCP_URL` | `http://127.0.0.1:8768/mcp` | The app's endpoint |
|
|
275
|
+
| `HIVE_AGENT_ID` | — | Agent id from the app; only for calls the app serves |
|
|
276
|
+
| `HIVE_AGENT_TOKEN` | — | Access token from the app; only for calls the app serves |
|
|
277
|
+
|
|
278
|
+
`HIVE_AGENT_ID` and `HIVE_AGENT_TOKEN` are **not** required to use the local store. They are
|
|
279
|
+
read straight from the environment and never written to disk by this package.
|
|
280
|
+
|
|
281
|
+
## Honest limits
|
|
282
|
+
|
|
283
|
+
Written down because a memory tool that oversells itself is worse than one that does less.
|
|
284
|
+
|
|
285
|
+
1. **The local store is not encrypted.** Plain SQLite. Anyone who can read your home directory
|
|
286
|
+
can read it. The app's vaults are the encrypted option, and `stats` reports
|
|
287
|
+
`"encrypted": false` instead of leaving you to find out.
|
|
288
|
+
2. **No semantic search**, as above.
|
|
289
|
+
3. **One writer at a time.** WAL mode lets the CLI confirm while a server is connected, but this
|
|
290
|
+
is a personal memory file, not a multi-host service. Do not point two machines at one
|
|
291
|
+
database over a network share.
|
|
292
|
+
4. **The board is coordination, not memory.** Messages are for hand-offs between agents; they
|
|
293
|
+
are not searchable with `recall`.
|
|
294
|
+
5. **The app is macOS only.** The local store is not — it is pure Python.
|
|
295
|
+
|
|
296
|
+
## Troubleshooting
|
|
297
|
+
|
|
298
|
+
**"no confirmed memory matched this query" with entries pending.** Correct behaviour. Confirm
|
|
299
|
+
them: `memory-hive-mcp-cli list`, then `approve`.
|
|
300
|
+
|
|
301
|
+
**A call fails with a message about `HIVE_AGENT_ID`.** The app answered and refused the
|
|
302
|
+
credential. The call did **not** fall back to the local store, on purpose. Register an agent in
|
|
303
|
+
the app and set both variables, or set `HIVE_MODE=local` if you meant to use the local file.
|
|
304
|
+
|
|
305
|
+
**"Hive app not reachable ... and HIVE_MODE=remote forbids the local fallback".** You asked for
|
|
306
|
+
the app only and it is not running. Start it, or drop `HIVE_MODE=remote`.
|
|
307
|
+
|
|
308
|
+
**Where did my memory go?** `memory-hive-mcp-cli stats` prints the database path. If `HIVE_MODE`
|
|
309
|
+
is `auto` and the app is running, the app has been answering — the local file is untouched by
|
|
310
|
+
design.
|
|
311
|
+
|
|
312
|
+
## Development
|
|
313
|
+
|
|
314
|
+
```bash
|
|
315
|
+
PYTHONPATH=src python -m unittest discover -s tests -v
|
|
316
|
+
```
|
|
317
|
+
|
|
318
|
+
The suite needs no running app and no network: it stands up its own fake backend where a real
|
|
319
|
+
response is required, and points `HIVE_DB` at a temporary file so a test never touches a real
|
|
320
|
+
memory database.
|
|
321
|
+
|
|
322
|
+
Layout:
|
|
323
|
+
|
|
324
|
+
```
|
|
325
|
+
src/hive_mcp/store.py the memory engine (SQLite, stdlib only)
|
|
326
|
+
src/hive_mcp/stdio.py the MCP server: routing, envelopes, the app bridge
|
|
327
|
+
src/hive_mcp/cli.py the human's half: review and confirm
|
|
328
|
+
src/hive_mcp/tools.json the tool surface snapshot for the offline path
|
|
329
|
+
```
|
|
330
|
+
|
|
331
|
+
## License
|
|
332
|
+
|
|
333
|
+
MIT. Your memory file is yours; nothing here phones home.
|
|
@@ -0,0 +1,309 @@
|
|
|
1
|
+
# Apiary Memory MCP
|
|
2
|
+
|
|
3
|
+
<!-- mcp-name: io.github.330899935/apiary-memory -->
|
|
4
|
+
|
|
5
|
+
> **Also known as:** Apiary Memory · 记忆蜂巢 · Memory Hive · 蜂巢 ·
|
|
6
|
+
> `io.github.330899935/apiary-memory` (MCP Registry name, from v0.2.0) ·
|
|
7
|
+
> `hive-memory` (Registry name up to v0.1.9) · `memory-hive-mcp` (PyPI package until the rename ships).
|
|
8
|
+
> Not affiliated with [hive.com](https://hive.com) (team collaboration) or the `hive-mcp` Claude Code plugin.
|
|
9
|
+
|
|
10
|
+
**Local-first personal memory for AI agents, over MCP.** Your agents can search what you have
|
|
11
|
+
confirmed and propose what they think you should keep; **only you can confirm it.** Everything
|
|
12
|
+
lives in one file on your own disk, and nothing is sent anywhere.
|
|
13
|
+
|
|
14
|
+
It runs on its own. A working SQLite memory store ships inside this package — no account, no
|
|
15
|
+
cloud, no daemon, no dependencies.
|
|
16
|
+
|
|
17
|
+
```bash
|
|
18
|
+
pip install memory-hive-mcp
|
|
19
|
+
```
|
|
20
|
+
|
|
21
|
+
**Desktop app, no terminal:** download [`memory-hive-mcp-0.1.9.mcpb`](https://github.com/330899935/apiary-memory/releases/download/v0.1.9/memory-hive-mcp-0.1.9.mcpb) and double-click it.
|
|
22
|
+
|
|
23
|
+
Measured on 2026-10-27, on one MacBook with the Hive running locally — script and raw output in [`bench/`](bench/):
|
|
24
|
+
|
|
25
|
+
| Measure | Result |
|
|
26
|
+
| --- | --- |
|
|
27
|
+
| Search latency | **p95 111 ms**, p50 76 ms (warm, includes HTTP) |
|
|
28
|
+
| Retrieval | correct entry in **top-5 for 97.5%** of 40 trials |
|
|
29
|
+
| Corpus | **1,303 entries · 2.39 M chars · 45 days**, plaintext only |
|
|
30
|
+
|
|
31
|
+
*Retrieval used each entry's own title as the query — an easy set. It shows the index is complete and ranks correctly on exact terms, not how well it answers a question asked in your own words. Run [`bench/bench_search.py`](bench/bench_search.py) on your own Hive.*
|
|
32
|
+
|
|
33
|
+
---
|
|
34
|
+
|
|
35
|
+
## What you get without the app
|
|
36
|
+
|
|
37
|
+
A complete, if modest, memory server. On a machine with nothing else installed:
|
|
38
|
+
|
|
39
|
+
| Call | What happens |
|
|
40
|
+
| --- | --- |
|
|
41
|
+
| `initialize` | answered locally, protocol `2024-11-05` |
|
|
42
|
+
| `tools/list` | answered locally, six tools |
|
|
43
|
+
| `propose` | **writes to a real database** on your disk |
|
|
44
|
+
| `recall` | **searches that database** and returns ranked entries |
|
|
45
|
+
| `get_context` | a digest of what you confirmed most recently |
|
|
46
|
+
| `manage` | read one entry, list what is pending, forget, restore |
|
|
47
|
+
| `say` / `board` | a shared message board for coordinating several agents |
|
|
48
|
+
|
|
49
|
+
Storage is SQLite at `~/.memory-hive/hive.db` by default. Search is lexical — character bigrams
|
|
50
|
+
for Chinese, words for Latin script — so it ranks well when the words overlap and finds nothing
|
|
51
|
+
when they do not. There are no embeddings and no vector index. Every response says
|
|
52
|
+
`"mode": "local"` so you can always tell which engine answered.
|
|
53
|
+
|
|
54
|
+
**Without the app, the database is plain SQLite: it is not encrypted.** That is stated here
|
|
55
|
+
rather than discovered later — see [Honest limits](#honest-limits).
|
|
56
|
+
|
|
57
|
+
## What the app adds
|
|
58
|
+
|
|
59
|
+
[Apiary Memory for macOS](https://apps.apple.com/app/id6806992983) is the same design with the
|
|
60
|
+
expensive parts built: eight vaults with four of them locked behind a PIN (Argon2id-wrapped key,
|
|
61
|
+
AES-256 at rest), a daily evolution log, multi-agent identity management, and a UI for reading
|
|
62
|
+
and confirming. It also answers MCP itself, with richer ranking.
|
|
63
|
+
|
|
64
|
+
If the app is running, this package uses it and leaves the local file alone. If it is not, the
|
|
65
|
+
local store answers. You do not have to choose up front:
|
|
66
|
+
|
|
67
|
+
| `HIVE_MODE` | Behaviour |
|
|
68
|
+
| --- | --- |
|
|
69
|
+
| `auto` (default) | Ask the app. Use the local store **only if the app is unreachable.** |
|
|
70
|
+
| `local` | Never touch the network. Always use the local store. |
|
|
71
|
+
| `remote` | Only the app. A connection failure is an error, not a reroute. |
|
|
72
|
+
|
|
73
|
+
One rule holds in every mode: **if the app is running and answers with an error, that error is
|
|
74
|
+
returned unchanged.** Falling back on a refusal would write your memory into a second database
|
|
75
|
+
while your client believed it had reached the app. Silent data forking is worse than an error
|
|
76
|
+
message, so a reply from the app is final. The fallback fires on a connection failure, and on
|
|
77
|
+
nothing else.
|
|
78
|
+
|
|
79
|
+
## Install
|
|
80
|
+
|
|
81
|
+
```bash
|
|
82
|
+
pip install memory-hive-mcp # the server: memory-hive-mcp
|
|
83
|
+
# the human's CLI: memory-hive-mcp-cli
|
|
84
|
+
```
|
|
85
|
+
|
|
86
|
+
Python 3.8 or newer. There is nothing else to install — the package uses only the standard
|
|
87
|
+
library, and `sqlite3` is part of it.
|
|
88
|
+
|
|
89
|
+
From a checkout:
|
|
90
|
+
|
|
91
|
+
```bash
|
|
92
|
+
git clone https://github.com/330899935/apiary-memory
|
|
93
|
+
cd apiary-memory
|
|
94
|
+
PYTHONPATH=src python -m hive_mcp.stdio
|
|
95
|
+
```
|
|
96
|
+
|
|
97
|
+
## Connect a client
|
|
98
|
+
|
|
99
|
+
Any MCP host that can run a stdio command will work. The minimum is the command; everything else
|
|
100
|
+
is optional.
|
|
101
|
+
|
|
102
|
+
**Claude Desktop** — `~/Library/Application Support/Claude/claude_desktop_config.json`:
|
|
103
|
+
|
|
104
|
+
```json
|
|
105
|
+
{
|
|
106
|
+
"mcpServers": {
|
|
107
|
+
"memory-hive": {
|
|
108
|
+
"command": "memory-hive-mcp",
|
|
109
|
+
"env": { "HIVE_MODE": "local" }
|
|
110
|
+
}
|
|
111
|
+
}
|
|
112
|
+
}
|
|
113
|
+
```
|
|
114
|
+
|
|
115
|
+
**Cursor** — `~/.cursor/mcp.json`, same `mcpServers` shape.
|
|
116
|
+
|
|
117
|
+
**Any host that only speaks HTTP** — start the app and point it at `http://127.0.0.1:8768/mcp`
|
|
118
|
+
with the headers `X-Hive-Agent` and `X-Hive-Token`. This package is the stdio path; the HTTP
|
|
119
|
+
endpoint belongs to the app.
|
|
120
|
+
|
|
121
|
+
## Prerequisite
|
|
122
|
+
|
|
123
|
+
**None, for the local store** — install the package and it works. A running app is a prerequisite
|
|
124
|
+
only for the app's own features: encrypted vaults, the evolution log, and multi-agent identity.
|
|
125
|
+
If you want those, install [the Hive app](https://apps.apple.com/app/id6806992983) (macOS,
|
|
126
|
+
version **1.4.5 or newer**) and leave `HIVE_MODE` at `auto`.
|
|
127
|
+
|
|
128
|
+
## The six tools
|
|
129
|
+
|
|
130
|
+
The six tools are the whole surface. They are the same six whether the app or the local store
|
|
131
|
+
answers; only the ranking behind them differs — and, for two inputs, whether they exist at all.
|
|
132
|
+
The app accepts precomputed embeddings (`recall.query_vector`, `propose.vector`); the local
|
|
133
|
+
store has no embedder, so `tools/list` omits those two wherever the local store can answer,
|
|
134
|
+
including under `auto` where the app might stop answering mid-session. A field that would be
|
|
135
|
+
accepted and silently ignored is not advertised at all.
|
|
136
|
+
|
|
137
|
+
The same rule covers wording. `recall.query` is described as the lexical ranker it is wherever
|
|
138
|
+
the local store can answer — words and character bigrams, no embeddings — and as the app's
|
|
139
|
+
index-based ranking with `query_vector` for meaning wherever only the app can answer. A
|
|
140
|
+
description is a claim about what will happen to the input, so it is narrowed with the inputs.
|
|
141
|
+
|
|
142
|
+
Defaults are not narrowed, they are shared. Omit `limit` and you get 10 from `recall` and 8 from
|
|
143
|
+
`get_context`; omit `compact` and bodies come back trimmed; omit `ttl` on `say` and the card
|
|
144
|
+
lives 24 hours. Both engines answer those numbers identically, because a default is a number in
|
|
145
|
+
the schema rather than a capability one engine lacks. `store.py` holds each one as a named
|
|
146
|
+
constant, and the tests compare the answer against the advertised number instead of a second
|
|
147
|
+
hand-written copy of it.
|
|
148
|
+
|
|
149
|
+
| Tool | Kind | What it is for |
|
|
150
|
+
| --- | --- | --- |
|
|
151
|
+
| `recall` | read | Search confirmed memory by topic. `{"query": "..."}` plus optional `limit`, `vaults`, `compact`. |
|
|
152
|
+
| `propose` | write | Submit a memory for the human to approve. Lands in a queue; changes no confirmed memory. |
|
|
153
|
+
| `manage` | read/write | Address one entry: `list_pending`, `get_by_id`, `forget`, `restore`, `discard_pending`. |
|
|
154
|
+
| `get_context` | read | A short digest of recent confirmed memory, to seed a conversation. |
|
|
155
|
+
| `say` | write | Post to the shared board so other agents see what was done or handed over. |
|
|
156
|
+
| `board` | read | Read the board, optionally only what arrived since your last cursor. |
|
|
157
|
+
|
|
158
|
+
Two properties are worth stating plainly, because they are the product:
|
|
159
|
+
|
|
160
|
+
- **`propose` succeeding means "queued", not "stored".** `recall` will not return it until a
|
|
161
|
+
human confirms it.
|
|
162
|
+
- **`manage` has no `approve` action.** There is no path from an agent to a confirmed memory.
|
|
163
|
+
That is not an omission; it is the design.
|
|
164
|
+
|
|
165
|
+
## Response shapes
|
|
166
|
+
|
|
167
|
+
`tools/list` is one surface for both engines, and an answer should be too: a model that has
|
|
168
|
+
to branch on which engine replied will eventually branch wrong. `board` therefore answers
|
|
169
|
+
the app's shape, field for field — `items`, `unread`, `count`, `cursor`, `reader` — and
|
|
170
|
+
`cursor` is the id of the newest card, the same kind of value the app returns.
|
|
171
|
+
|
|
172
|
+
The cursor is what a poll carries *across* an engine switch, which is why it gets the space.
|
|
173
|
+
Under `auto` one call can be answered by the app and the next by the local store. Before 0.1.9
|
|
174
|
+
the local engine answered a float epoch timestamp and read the app's card id with `float()`,
|
|
175
|
+
so a cursor fetched from the app came back as a full re-read of the board and nothing said
|
|
176
|
+
so; a float sent the other way matched no card id, and the app answered with its newest
|
|
177
|
+
`limit`, equally quietly.
|
|
178
|
+
|
|
179
|
+
A cursor can still fail to resolve: its card expired (cards live 24 hours by default), or it
|
|
180
|
+
was issued by the other engine, whose ids are `CB…` in the app and `msg_…` here. Neither
|
|
181
|
+
engine can resume from that, so both return the newest window — but this one also sets
|
|
182
|
+
`since_unknown` and a `note` saying which of those happened. `unread` counts every live card
|
|
183
|
+
after the cursor and `count` counts what the answer carries, so `unread > count` means the
|
|
184
|
+
page did not cover everything waiting, and the `note` says so in words.
|
|
185
|
+
|
|
186
|
+
Three shapes are **not** unified, because the engines really do hold different data:
|
|
187
|
+
|
|
188
|
+
- a `recall` entry names its vault `vault` from the local store and `wh` from the app;
|
|
189
|
+
- `get_context` returns one flat `digest` locally and one bucket per warehouse (`context`,
|
|
190
|
+
keyed by warehouse name) from the app;
|
|
191
|
+
- `roster` entries share `agent_id`; locally they also carry `messages` and `last_seen`,
|
|
192
|
+
while the app adds its registry card — display name, seat, leases, current task — which a
|
|
193
|
+
SQLite file has no way to know.
|
|
194
|
+
|
|
195
|
+
## Confirming what agents propose
|
|
196
|
+
|
|
197
|
+
Agents propose. You confirm. The confirmation lives in a terminal, in a command you have to
|
|
198
|
+
type — deliberately not a tool an agent can call.
|
|
199
|
+
|
|
200
|
+
```bash
|
|
201
|
+
memory-hive-mcp-cli list # what is waiting
|
|
202
|
+
memory-hive-mcp-cli show <id> # read one in full
|
|
203
|
+
memory-hive-mcp-cli approve <id> # confirm it, or --all
|
|
204
|
+
memory-hive-mcp-cli reject <id> # drop it for good
|
|
205
|
+
memory-hive-mcp-cli search "deploy key" # search confirmed memory
|
|
206
|
+
memory-hive-mcp-cli forget <id> # soft delete, reversible
|
|
207
|
+
memory-hive-mcp-cli restore <id> # undo forget
|
|
208
|
+
memory-hive-mcp-cli stats # counts and the database path
|
|
209
|
+
memory-hive-mcp-cli export --out dump.json
|
|
210
|
+
memory-hive-mcp-cli path
|
|
211
|
+
```
|
|
212
|
+
|
|
213
|
+
`--db <path>` on any command points at a different database. `HIVE_DB` does the same from the
|
|
214
|
+
environment.
|
|
215
|
+
|
|
216
|
+
## Where your memory lives
|
|
217
|
+
|
|
218
|
+
`~/.memory-hive/hive.db`, unless `HIVE_DB` says otherwise. One SQLite file; copy it to back it
|
|
219
|
+
up, delete it to start over.
|
|
220
|
+
|
|
221
|
+
The board keeps its messages in the same file, so there is only ever one thing to back up. Cards
|
|
222
|
+
expire: 24 hours unless `ttl` says otherwise, clamped to between 60 seconds and 30 days, which
|
|
223
|
+
is what the app does too — a board that grows forever stops being a board.
|
|
224
|
+
|
|
225
|
+
## How search works, and what it does not do
|
|
226
|
+
|
|
227
|
+
Chinese text is indexed as single characters plus overlapping character bigrams; Latin text as
|
|
228
|
+
words. A query is scored by how much of it the entry covers, then nudged by recency — a 10%
|
|
229
|
+
multiplier, enough to break ties and not enough to overturn relevance.
|
|
230
|
+
|
|
231
|
+
This is a ranker, not an understander.
|
|
232
|
+
|
|
233
|
+
- **Scores are relative.** `0.8` means "closer than `0.4` for this query", not "80% confident".
|
|
234
|
+
They are not comparable across queries.
|
|
235
|
+
- **No semantic recall.** Ask about "money" and an entry that only says "invoice" will not be
|
|
236
|
+
found. Related-word coverage needs an embedder; this package has no dependencies and does not
|
|
237
|
+
pretend otherwise.
|
|
238
|
+
- **The ranking is the one the description states.** `tools/list` says "words and character
|
|
239
|
+
bigrams ... no embeddings" here, and that sentence is true of this engine. It is narrowed per
|
|
240
|
+
engine like the embedding inputs are, because the app's answer to the same question differs.
|
|
241
|
+
- **A miss is reported as a miss.** When nothing matches, `recall` says so and mentions how many
|
|
242
|
+
entries are awaiting confirmation, rather than returning a plausible neighbour.
|
|
243
|
+
|
|
244
|
+
## Configuration
|
|
245
|
+
|
|
246
|
+
| Variable | Default | Purpose |
|
|
247
|
+
| --- | --- | --- |
|
|
248
|
+
| `HIVE_MODE` | `auto` | `auto`, `local`, or `remote` — see the table above |
|
|
249
|
+
| `HIVE_DB` | `~/.memory-hive/hive.db` | Where the local database lives |
|
|
250
|
+
| `HIVE_MCP_URL` | `http://127.0.0.1:8768/mcp` | The app's endpoint |
|
|
251
|
+
| `HIVE_AGENT_ID` | — | Agent id from the app; only for calls the app serves |
|
|
252
|
+
| `HIVE_AGENT_TOKEN` | — | Access token from the app; only for calls the app serves |
|
|
253
|
+
|
|
254
|
+
`HIVE_AGENT_ID` and `HIVE_AGENT_TOKEN` are **not** required to use the local store. They are
|
|
255
|
+
read straight from the environment and never written to disk by this package.
|
|
256
|
+
|
|
257
|
+
## Honest limits
|
|
258
|
+
|
|
259
|
+
Written down because a memory tool that oversells itself is worse than one that does less.
|
|
260
|
+
|
|
261
|
+
1. **The local store is not encrypted.** Plain SQLite. Anyone who can read your home directory
|
|
262
|
+
can read it. The app's vaults are the encrypted option, and `stats` reports
|
|
263
|
+
`"encrypted": false` instead of leaving you to find out.
|
|
264
|
+
2. **No semantic search**, as above.
|
|
265
|
+
3. **One writer at a time.** WAL mode lets the CLI confirm while a server is connected, but this
|
|
266
|
+
is a personal memory file, not a multi-host service. Do not point two machines at one
|
|
267
|
+
database over a network share.
|
|
268
|
+
4. **The board is coordination, not memory.** Messages are for hand-offs between agents; they
|
|
269
|
+
are not searchable with `recall`.
|
|
270
|
+
5. **The app is macOS only.** The local store is not — it is pure Python.
|
|
271
|
+
|
|
272
|
+
## Troubleshooting
|
|
273
|
+
|
|
274
|
+
**"no confirmed memory matched this query" with entries pending.** Correct behaviour. Confirm
|
|
275
|
+
them: `memory-hive-mcp-cli list`, then `approve`.
|
|
276
|
+
|
|
277
|
+
**A call fails with a message about `HIVE_AGENT_ID`.** The app answered and refused the
|
|
278
|
+
credential. The call did **not** fall back to the local store, on purpose. Register an agent in
|
|
279
|
+
the app and set both variables, or set `HIVE_MODE=local` if you meant to use the local file.
|
|
280
|
+
|
|
281
|
+
**"Hive app not reachable ... and HIVE_MODE=remote forbids the local fallback".** You asked for
|
|
282
|
+
the app only and it is not running. Start it, or drop `HIVE_MODE=remote`.
|
|
283
|
+
|
|
284
|
+
**Where did my memory go?** `memory-hive-mcp-cli stats` prints the database path. If `HIVE_MODE`
|
|
285
|
+
is `auto` and the app is running, the app has been answering — the local file is untouched by
|
|
286
|
+
design.
|
|
287
|
+
|
|
288
|
+
## Development
|
|
289
|
+
|
|
290
|
+
```bash
|
|
291
|
+
PYTHONPATH=src python -m unittest discover -s tests -v
|
|
292
|
+
```
|
|
293
|
+
|
|
294
|
+
The suite needs no running app and no network: it stands up its own fake backend where a real
|
|
295
|
+
response is required, and points `HIVE_DB` at a temporary file so a test never touches a real
|
|
296
|
+
memory database.
|
|
297
|
+
|
|
298
|
+
Layout:
|
|
299
|
+
|
|
300
|
+
```
|
|
301
|
+
src/hive_mcp/store.py the memory engine (SQLite, stdlib only)
|
|
302
|
+
src/hive_mcp/stdio.py the MCP server: routing, envelopes, the app bridge
|
|
303
|
+
src/hive_mcp/cli.py the human's half: review and confirm
|
|
304
|
+
src/hive_mcp/tools.json the tool surface snapshot for the offline path
|
|
305
|
+
```
|
|
306
|
+
|
|
307
|
+
## License
|
|
308
|
+
|
|
309
|
+
MIT. Your memory file is yours; nothing here phones home.
|