plugmem 0.4.0 → 0.5.0
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/README.md +408 -198
- package/index.d.ts +32 -4
- package/package.json +7 -7
package/README.md
CHANGED
|
@@ -1,20 +1,25 @@
|
|
|
1
|
-
# plugmem
|
|
1
|
+
# plugmem
|
|
2
2
|
|
|
3
3
|
> ⚠️ Experimental. plugmem is mostly an AI-built experiment — written with
|
|
4
4
|
> the help of a small local model (Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf) and various
|
|
5
5
|
> Claude models, in roughly equal measure. Expect non-professional design
|
|
6
6
|
> choices, rough edges, broken behavior, or mistakes. Use it at your own risk.
|
|
7
7
|
|
|
8
|
-
|
|
9
|
-
|
|
10
|
-
[`plugmem-host`](https://docs.rs/plugmem-host/latest) **in the Node process**
|
|
11
|
-
(real mmap, file locking, cross-process MVCC — the whole engine, unchanged) and
|
|
12
|
-
exposes it to **JavaScript / TypeScript** as a `Plugmem` class. It is published
|
|
13
|
-
to npm as **`plugmem`**.
|
|
8
|
+
A memory database for LLM agents, embedded in your Node process. It stores short
|
|
9
|
+
facts and answers a query with a ranked block of text ready to put in a prompt.
|
|
14
10
|
|
|
15
|
-
|
|
16
|
-
|
|
17
|
-
|
|
11
|
+
One file on disk, no server, no daemon. It links into your process the way
|
|
12
|
+
SQLite does: the engine is [`plugmem-host`](https://docs.rs/plugmem-host/latest)
|
|
13
|
+
compiled to a native addon through [napi-rs](https://napi.rs), so there is no
|
|
14
|
+
WebAssembly copy of the file in RAM and no 4 GiB ceiling. Runs on Node, Deno and
|
|
15
|
+
Bun.
|
|
16
|
+
|
|
17
|
+
**Contents:** [Install](#install) · [Quick start](#quick-start) ·
|
|
18
|
+
[What it stores](#what-it-stores) · [Two clocks](#two-clocks) ·
|
|
19
|
+
[How recall works](#how-recall-works) · [API](#api) · [Errors](#errors) ·
|
|
20
|
+
[Configuration](#configuration-and-embeddings) ·
|
|
21
|
+
[Async](#async-and-the-event-loop) · [Many memories](#many-memories-in-one-directory) ·
|
|
22
|
+
[What it is not for](#what-it-is-not-for)
|
|
18
23
|
|
|
19
24
|
## Install
|
|
20
25
|
|
|
@@ -22,115 +27,226 @@ Rust library. It loads in **Node, and any N-API host (Deno, Bun)**.
|
|
|
22
27
|
$ npm install plugmem
|
|
23
28
|
```
|
|
24
29
|
|
|
25
|
-
|
|
26
|
-
|
|
27
|
-
|
|
28
|
-
|
|
30
|
+
That pulls a meta package which, through `optionalDependencies`, installs only
|
|
31
|
+
the prebuilt binary for your platform: one of `plugmem-{linux-x64-gnu,
|
|
32
|
+
linux-arm64-gnu, darwin-x64, darwin-arm64, win32-x64-msvc, win32-arm64-msvc}`.
|
|
33
|
+
No toolchain, no build step.
|
|
34
|
+
|
|
35
|
+
## Quick start
|
|
36
|
+
|
|
37
|
+
```typescript
|
|
38
|
+
import { Plugmem } from "plugmem";
|
|
39
|
+
|
|
40
|
+
const db = await Plugmem.open("agent.plugmem");
|
|
41
|
+
|
|
42
|
+
await db.remember({ text: "the user prefers tokio", entity: "user", tags: ["pref"] });
|
|
43
|
+
await db.remember({ text: "the release ships on friday", entity: "release" });
|
|
44
|
+
|
|
45
|
+
const res = await db.recall({ query: "which runtime?", k: 5 });
|
|
46
|
+
console.log(res.rendered); // paste this into the prompt
|
|
47
|
+
// - [f0] user: the user prefers tokio (2026-08; active)
|
|
48
|
+
|
|
49
|
+
db.close();
|
|
50
|
+
```
|
|
51
|
+
|
|
52
|
+
`Plugmem.open` is a static method, not a constructor, because opening replays a
|
|
53
|
+
journal and maps a snapshot — work proportional to the file — and a JavaScript
|
|
54
|
+
constructor has no way to hand that to a worker thread. Everything is typed:
|
|
55
|
+
`index.d.ts` is generated from the Rust, so a TypeScript host gets real
|
|
56
|
+
autocomplete on arguments and results.
|
|
57
|
+
|
|
58
|
+
## What it stores
|
|
59
|
+
|
|
60
|
+
One **fact** is one statement. It carries:
|
|
61
|
+
|
|
62
|
+
| Field | Meaning |
|
|
63
|
+
|---|---|
|
|
64
|
+
| `text` | the statement itself, and what lexical search indexes |
|
|
65
|
+
| `entity` | the subject, by name — created on first mention, shared across facts |
|
|
66
|
+
| `tags` | filters, not ranking: a query asking for a tag requires it |
|
|
67
|
+
| `metadata` | an opaque `Record<string,string>`. The engine stores and returns it and never looks inside — use it for a URI to the real payload elsewhere, a mime type, an external key |
|
|
68
|
+
| `vector` | an optional embedding; supply your own, or let a configured embedder produce it |
|
|
69
|
+
| `validFrom` | when the statement became true |
|
|
70
|
+
|
|
71
|
+
Entities are joined by **typed edges**: `link({ src: "ann", rel: "hires", dst: "bob" })`.
|
|
72
|
+
An edge can name the fact it follows from (`provenance`), so a later reader can
|
|
73
|
+
answer "why does the memory think this" instead of trusting a bare relationship.
|
|
29
74
|
|
|
30
|
-
|
|
75
|
+
Facts are never rewritten in place. `revise` closes the old one and chains a
|
|
76
|
+
successor; `forget` tombstones a fact and the next `maintain` erases it from
|
|
77
|
+
disk. `unlink` closes an edge the same way `revise` closes a fact.
|
|
31
78
|
|
|
32
|
-
|
|
79
|
+
## Two clocks
|
|
33
80
|
|
|
34
|
-
|
|
35
|
-
|
|
36
|
-
| **writing JavaScript / TypeScript for Node** | **`plugmem-napi`** (this, npm `plugmem`) | The engine *in your Node process*, native speed, typed for TS. |
|
|
37
|
-
| **writing Rust** | [`plugmem-host`](https://docs.rs/plugmem-host/latest) | The engine in your process, like linking SQLite. |
|
|
38
|
-
| **an agent, or another language** (Python, Go…) | [`plugmem-mcp`](https://docs.rs/plugmem-mcp/latest) | A long-lived stdio JSON-RPC sidecar; language-independent. |
|
|
39
|
-
| a person at a **terminal / script** | [`plugmem-cli`](https://docs.rs/plugmem-cli/latest) | The human door. |
|
|
81
|
+
This is the part worth reading, because it is the one thing that behaves
|
|
82
|
+
differently from every other store.
|
|
40
83
|
|
|
41
|
-
|
|
42
|
-
the CLI.**
|
|
84
|
+
Every fact carries **two** timestamps, not one:
|
|
43
85
|
|
|
44
|
-
|
|
86
|
+
- **`validFrom` / `validTo`** — when the statement was true.
|
|
87
|
+
- **`recordedAt`** — when the memory learned it. Set by the engine, never by you.
|
|
45
88
|
|
|
46
|
-
|
|
47
|
-
|
|
48
|
-
|
|
49
|
-
|
|
89
|
+
They are different questions, and one timestamp cannot hold both.
|
|
90
|
+
|
|
91
|
+
Unlike the Rust library, this binding reads the system clock on every call — you
|
|
92
|
+
never pass `now` — so `recordedAt` is always the moment of the write:
|
|
50
93
|
|
|
51
94
|
```typescript
|
|
52
|
-
|
|
95
|
+
await db.remember({ text: "lives in Moscow", entity: "kim" });
|
|
96
|
+
const between = Date.now();
|
|
97
|
+
await db.revise(0, { text: "lives in Berlin", entity: "kim" });
|
|
53
98
|
|
|
54
|
-
|
|
99
|
+
(await db.recall({ entities: ["kim"] })).rendered;
|
|
100
|
+
// - [f1] kim: lives in Berlin (2026-08; active)
|
|
55
101
|
|
|
56
|
-
|
|
57
|
-
|
|
58
|
-
|
|
59
|
-
tags: ["pref"],
|
|
60
|
-
links: [{ rel: "works_at", entity: "acme" }],
|
|
61
|
-
metadata: { source: "chat", uri: "s3://bucket/note.txt" }, // opaque; a pointer
|
|
62
|
-
});
|
|
63
|
-
out.id; // number
|
|
64
|
-
out.similar; // Similar[] — the engine surfaces conflicts, you decide
|
|
102
|
+
(await db.recall({ entities: ["kim"], asOf: between })).rendered;
|
|
103
|
+
// - [f0] kim: lives in Moscow (2026-08 → 2026-08; closed)
|
|
104
|
+
```
|
|
65
105
|
|
|
66
|
-
|
|
67
|
-
|
|
68
|
-
res.facts; // RecalledFact[] — { id, score, entity, recordedAt, … }
|
|
106
|
+
`revise` closed the first fact's interval rather than deleting it, which is why
|
|
107
|
+
the second query has something to answer with.
|
|
69
108
|
|
|
70
|
-
|
|
71
|
-
|
|
109
|
+
`asOf` moves **both** clocks: a fact answers only if it was valid at that
|
|
110
|
+
instant *and* had already been recorded by then. The second half is the one
|
|
111
|
+
people trip over — an `asOf` earlier than a fact's `recordedAt` sees nothing,
|
|
112
|
+
because the memory genuinely knew nothing then. Answering with today's knowledge
|
|
113
|
+
would be the wrong answer to "what did I hold".
|
|
72
114
|
|
|
73
|
-
|
|
74
|
-
|
|
75
|
-
|
|
115
|
+
`validFrom` is the other half: a statement that became true before you heard of
|
|
116
|
+
it. Recording today that someone moved a week ago closes the previous interval a
|
|
117
|
+
week ago rather than now, so a query as of three days back finds **neither** —
|
|
118
|
+
the old fact had stopped being true, and the new one was not yet known. That is
|
|
119
|
+
not a hole in the model; it is the honest answer for that instant, and it is
|
|
120
|
+
what a single timestamp cannot express.
|
|
76
121
|
|
|
77
|
-
|
|
78
|
-
|
|
122
|
+
Two more queries over the same axes:
|
|
123
|
+
|
|
124
|
+
```typescript
|
|
125
|
+
await db.recall({ range: [from, to] }); // what did I record in this window
|
|
126
|
+
await db.recall({ query: "kim", closed: true }); // include closed revisions
|
|
79
127
|
```
|
|
80
128
|
|
|
81
|
-
|
|
129
|
+
Use `revise` when something changed and `forget` only when a fact was simply
|
|
130
|
+
wrong: `forget` destroys the "what was true then" answer, `revise` keeps it.
|
|
131
|
+
|
|
132
|
+
Edges are temporal too, so `asOf` walks the graph as it stood then — through
|
|
133
|
+
relationships that have since been unlinked.
|
|
134
|
+
|
|
135
|
+
## How recall works
|
|
82
136
|
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
137
|
+
Not a vector lookup. Four sources run and are fused by
|
|
138
|
+
[reciprocal-rank fusion](https://plg.uwaterloo.ca/~gvcormac/cormacksigir09-rrf.pdf)
|
|
139
|
+
with a recency boost; tags filter and are not a source:
|
|
140
|
+
|
|
141
|
+
| Source | What it finds |
|
|
142
|
+
|---|---|
|
|
143
|
+
| **Lexical** — [BM25](https://en.wikipedia.org/wiki/Okapi_BM25) over a Unicode ([UAX #29](https://unicode.org/reports/tr29/)) tokenizer | exact terms, keyword overlap |
|
|
144
|
+
| **Semantic** — int8-quantized cosine, a flat scan below a threshold and an [HNSW](https://arxiv.org/abs/1603.09320) graph above | meaning, nearest neighbours |
|
|
145
|
+
| **Graph** — typed edges walked from the query's anchor entities | relational knowledge |
|
|
146
|
+
| **Temporal** — range scans over the `recordedAt` index, plus the validity test | "what was true then", time windows |
|
|
147
|
+
|
|
148
|
+
The sources compose. A query with no `query` string still answers from tags,
|
|
149
|
+
entities and time. **Without an embedder the system is complete** — the other
|
|
150
|
+
three sources need no model, no network and no API key.
|
|
151
|
+
|
|
152
|
+
The result carries both a `rendered` block, selected greedily under a token
|
|
153
|
+
budget and ready to paste, and the structured `facts`/`edges` behind it:
|
|
89
154
|
|
|
90
155
|
```typescript
|
|
91
|
-
const
|
|
156
|
+
const res = await db.recall({
|
|
157
|
+
query: "release plans",
|
|
158
|
+
entities: ["ann"], // graph anchors
|
|
159
|
+
tags: ["work"], // filter: a fact must carry all of these
|
|
160
|
+
k: 10, // cap the number of facts
|
|
161
|
+
tokenBudget: 400, // cap the size of the block — your context budget
|
|
162
|
+
});
|
|
163
|
+
|
|
164
|
+
res.rendered; // string, prompt-ready
|
|
165
|
+
res.facts; // { id, score, entity, recordedAt, validFrom, validTo, sources }[]
|
|
166
|
+
res.edges; // { src, rel, dst, provenance }[] — what the graph walked
|
|
167
|
+
res.truncated; // true if selection stopped at k or the budget with more left
|
|
92
168
|
```
|
|
93
169
|
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
path = "/path/to/memory.plugmem" # optional example
|
|
170
|
+
`remember` returns the new fact's id **plus any live facts it may duplicate or
|
|
171
|
+
contradict**. The engine never merges or deletes on its own — it surfaces the
|
|
172
|
+
tension and your code decides:
|
|
98
173
|
|
|
99
|
-
|
|
100
|
-
|
|
174
|
+
```typescript
|
|
175
|
+
const out = await db.remember({ text: "the user prefers async-std", entity: "user" });
|
|
176
|
+
for (const s of out.similar) {
|
|
177
|
+
// s.id, s.score, s.reason: "LexicalOverlap" | "VectorCosine"
|
|
178
|
+
// → revise it, forget it, or keep both. Your call.
|
|
179
|
+
}
|
|
180
|
+
```
|
|
101
181
|
|
|
102
|
-
|
|
103
|
-
|
|
104
|
-
|
|
105
|
-
|
|
182
|
+
## API
|
|
183
|
+
|
|
184
|
+
Every method wraps the identically-named verb of the Rust `Database`; this layer
|
|
185
|
+
only moves arguments and results across the boundary.
|
|
186
|
+
|
|
187
|
+
**Writing** — all return promises:
|
|
188
|
+
|
|
189
|
+
| Method | Does |
|
|
190
|
+
|---|---|
|
|
191
|
+
| `remember(args)` | store a fact; resolves with its id and similar facts |
|
|
192
|
+
| `rememberMany(args[])` | store a batch: one embedding round-trip, one journal sync |
|
|
193
|
+
| `revise(id, args)` | close a fact and record its successor |
|
|
194
|
+
| `forget(id)` | tombstone a fact; resolves with whether it was live |
|
|
195
|
+
| `link(args)` | upsert a typed edge, optionally with `provenance` |
|
|
196
|
+
| `unlink(args)` | close the current edge; resolves with whether one was open |
|
|
197
|
+
|
|
198
|
+
**Reading** — synchronous ones touch mapped memory and return in microseconds:
|
|
199
|
+
|
|
200
|
+
| Method | Does |
|
|
201
|
+
|---|---|
|
|
202
|
+
| `recall(args?)` | ranked, fused, token-budgeted result (async) |
|
|
203
|
+
| `get(id)` | one fact's full card, or `null` (sync) |
|
|
204
|
+
| `tagsOf(id)` | that fact's tags (sync) |
|
|
205
|
+
| `stats()` | engine size counters (sync) |
|
|
206
|
+
| `path()` | the file this handle resolved to (sync) |
|
|
207
|
+
| `export()` | every open fact as one array (async, unbounded — see below) |
|
|
208
|
+
| `exportPage(cursor?)` | the same data in bounded pages of 128 (async) |
|
|
209
|
+
|
|
210
|
+
**Upkeep** — all async, all on a worker thread:
|
|
211
|
+
|
|
212
|
+
| Method | Does |
|
|
213
|
+
|---|---|
|
|
214
|
+
| `maintain(mode?)` | `"auto"` (default), `"compact"`, `"reindex-text"`, `"optimize-vectors"`, `"full"`. No mode ever drops a revision or an edge version |
|
|
215
|
+
| `checkpoint()` | flush the journal into a fresh snapshot |
|
|
216
|
+
| `verify()` | full content-integrity sweep; rejects on the first inconsistency |
|
|
217
|
+
|
|
218
|
+
**Read-only handles** (`{ readOnly: true }`) observe another process's writer
|
|
219
|
+
over a published snapshot. The read verbs answer, the write verbs throw, and two
|
|
220
|
+
more appear: `generation()` (the pinned snapshot number) and `refresh()` (adopt
|
|
221
|
+
the writer's latest checkpoint, returning whether a newer one existed).
|
|
222
|
+
|
|
223
|
+
`close()` releases the file and its lock; every verb afterwards throws, and
|
|
224
|
+
calling it twice is a no-op.
|
|
225
|
+
|
|
226
|
+
### Bringing your own embedding
|
|
227
|
+
|
|
228
|
+
`remember`, `revise`, `rememberMany` and `recall` all take an optional `vector`
|
|
229
|
+
whose length must equal the configured `dim`. Given one, it **replaces** the
|
|
230
|
+
embedder for that call — nothing is sent to the provider:
|
|
231
|
+
|
|
232
|
+
```typescript
|
|
233
|
+
const own = await myEmbedder(text);
|
|
234
|
+
await db.remember({ text, vector: own });
|
|
235
|
+
const res = await db.recall({ query: text, vector: own });
|
|
106
236
|
```
|
|
107
237
|
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
111
|
-
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
|
|
116
|
-
|
|
117
|
-
|
|
118
|
-
|
|
119
|
-
|
|
120
|
-
Every method here is the identically-named `plugmem-host` `Database` verb; the
|
|
121
|
-
engine logic is entirely the host's.
|
|
122
|
-
|
|
123
|
-
**Writer** (default): `remember`, `rememberMany`, `recall`, `revise(id, args)`,
|
|
124
|
-
`forget(id)`, `link`, `unlink`, `get(id)`, `tagsOf(id)`, `stats`, `export`,
|
|
125
|
-
`exportPage(cursor?)`, `verify`, and the async maintenance verbs below.
|
|
126
|
-
**Read-only** (`{ readOnly: true }`, observing another process's writer):
|
|
127
|
-
`recall`, `get`, `tagsOf`, `stats`, `export`, `exportPage(cursor?)`, `verify`,
|
|
128
|
-
plus `generation()` (the pinned snapshot generation) and `refresh()` (advance
|
|
129
|
-
to the writer's latest checkpoint); the write verbs throw.
|
|
130
|
-
**Both**: `path()` — the file the handle resolved to, which is the only way to
|
|
131
|
-
learn it when the constructor was given no path.
|
|
132
|
-
|
|
133
|
-
### Errors
|
|
238
|
+
Use it for vectors you already have, for a model that is not an OpenAI-shaped
|
|
239
|
+
HTTP endpoint, or for a deterministic test with no network.
|
|
240
|
+
|
|
241
|
+
### What the Rust library has and this does not
|
|
242
|
+
|
|
243
|
+
| Not here | Why |
|
|
244
|
+
|---|---|
|
|
245
|
+
| `recover` | salvaging a damaged file is a path-level operation on the host's disk — the CLI's job |
|
|
246
|
+
| `import` | the JSONL file format lives in [`plugmem-cli`](https://docs.rs/plugmem-cli/latest). A Node program holding records already uses `rememberMany` |
|
|
247
|
+
| `scrub` | byte-level container integrity, likewise CLI-side; `verify()` covers content |
|
|
248
|
+
|
|
249
|
+
## Errors
|
|
134
250
|
|
|
135
251
|
Every failure plugmem itself decides carries a stable `code`, so a program
|
|
136
252
|
branches on it instead of on wording:
|
|
@@ -149,153 +265,247 @@ try {
|
|
|
149
265
|
`PLUGMEM_INVALID_NAME` from an argument that was refused; `PLUGMEM_CLOSED`,
|
|
150
266
|
`PLUGMEM_READ_ONLY`, `PLUGMEM_WRITER_ONLY` and `PLUGMEM_BUSY` from calling a
|
|
151
267
|
verb the handle cannot serve; `PLUGMEM_ENGINE` from the engine itself, carrying
|
|
152
|
-
|
|
268
|
+
its own message.
|
|
153
269
|
|
|
154
270
|
The code is there whether the verb threw or the promise rejected — the two are
|
|
155
271
|
the same contract, so nothing has to be handled twice.
|
|
156
272
|
|
|
157
|
-
An argument that
|
|
158
|
-
be exactly `[from, to]`, and `range`, `asOf` and `validFrom` must each be a
|
|
159
|
-
finite, non-negative instant. Silently ignoring one
|
|
160
|
-
without it
|
|
273
|
+
An argument that would shape an answer is refused rather than dropped: `range`
|
|
274
|
+
must be exactly `[from, to]`, and `range`, `asOf` and `validFrom` must each be a
|
|
275
|
+
finite, non-negative instant. Silently ignoring one produces an answer computed
|
|
276
|
+
without it, indistinguishable from a correct one.
|
|
161
277
|
|
|
162
|
-
|
|
278
|
+
## Configuration and embeddings
|
|
163
279
|
|
|
164
|
-
|
|
165
|
-
|
|
166
|
-
|
|
167
|
-
|
|
168
|
-
|---|---|
|
|
169
|
-
| `recover` | salvaging a damaged file is a path-level operation on the disk the process is running on — [`plugmem-cli recover`](https://docs.rs/plugmem-cli/latest)'s job, like `import` and `scrub`. |
|
|
170
|
-
| `remember_many` | Exposed as async `rememberMany(items)`. It writes a batch with one embedding round-trip and resolves with outcomes in input order. |
|
|
171
|
-
| `export_each` | Exposed as pull-based `exportPage(cursor?)`. Each Promise returns at most 128 facts and releases the native read lock before JS processes them; this gives Node backpressure without a cross-thread callback. |
|
|
172
|
-
| `tags_of` | Exposed as synchronous `tagsOf(id)`, returning one fact's tags or an empty array. |
|
|
280
|
+
Settings resolve from an explicit `config` path, then `$PLUGMEM_CONFIG`, then
|
|
281
|
+
the platform config directory, then defaults. The database path resolves from an
|
|
282
|
+
explicit argument, then `$PLUGMEM_DB`, then `[database].path`, then the platform
|
|
283
|
+
data directory.
|
|
173
284
|
|
|
174
|
-
|
|
175
|
-
|
|
176
|
-
|
|
285
|
+
```typescript
|
|
286
|
+
const db = await Plugmem.open(undefined, { config: "./plugmem.toml" });
|
|
287
|
+
```
|
|
177
288
|
|
|
178
|
-
|
|
289
|
+
```toml
|
|
290
|
+
# plugmem.toml
|
|
291
|
+
[database]
|
|
292
|
+
path = "/path/to/memory.plugmem"
|
|
179
293
|
|
|
180
|
-
|
|
294
|
+
[engine]
|
|
295
|
+
dim = 768 # embedding size (0 = vectors off)
|
|
181
296
|
|
|
182
|
-
|
|
183
|
-
|
|
297
|
+
[embedder] # optional — omit for lexical/tag/graph/time only
|
|
298
|
+
kind = "ollama" # or openai / lmstudio / vllm / llamacpp
|
|
299
|
+
url = "http://localhost:11434/v1/embeddings"
|
|
300
|
+
model = "nomic-embed-text"
|
|
184
301
|
|
|
185
|
-
|
|
186
|
-
|
|
302
|
+
[maintenance]
|
|
303
|
+
fsync = "each_op" # or "on_snapshot": faster, loses the journal tail on an OS crash
|
|
304
|
+
```
|
|
187
305
|
|
|
188
|
-
|
|
306
|
+
With an `[embedder]`, a text-only `remember`/`recall` embeds automatically, and
|
|
307
|
+
the provider's HTTP call happens outside the engine lock. The `dim` open option
|
|
308
|
+
sets the embedding size when there is no config; if the config built an
|
|
309
|
+
embedder, its dimension governs and `dim` must agree.
|
|
189
310
|
|
|
190
|
-
|
|
191
|
-
|
|
192
|
-
|
|
311
|
+
A read-only handle cannot embed inside the engine — writing into a zero-copy
|
|
312
|
+
mapping is exactly what read-only exists to avoid — so this binding embeds the
|
|
313
|
+
query itself before the read. A text `recall` reaches the vector source in both
|
|
314
|
+
modes.
|
|
193
315
|
|
|
194
|
-
|
|
195
|
-
|
|
196
|
-
await ws.describe("chat-42", { description: "release planning", owner: "ann" });
|
|
197
|
-
const hits: DbEntry[] = await ws.find("release planning"); // → [{ db: "chat-42", … }]
|
|
198
|
-
```
|
|
316
|
+
The [full settings reference](https://github.com/m62624/plugmem/blob/main/crates/plugmem-host/SETTINGS.md)
|
|
317
|
+
lists every field and the OS-specific paths.
|
|
199
318
|
|
|
200
|
-
|
|
201
|
-
to exactly one file inside the directory — traversal is not filtered out, it is
|
|
202
|
-
unconstructible. `open(name, false)` refuses a name that does not exist yet,
|
|
203
|
-
which is what a read should do so a typo is diagnosed rather than answered with
|
|
204
|
-
an empty result.
|
|
319
|
+
## Async and the event loop
|
|
205
320
|
|
|
206
|
-
|
|
207
|
-
|
|
208
|
-
process that
|
|
209
|
-
|
|
321
|
+
Node runs all JavaScript on one thread, so a native call that waits on an
|
|
322
|
+
embedder's HTTP round trip or an fsync would freeze every timer, socket and
|
|
323
|
+
callback in the process. Anything that can do that runs on a libuv worker and
|
|
324
|
+
returns a promise instead.
|
|
210
325
|
|
|
211
|
-
|
|
212
|
-
|
|
213
|
-
|
|
214
|
-
|
|
215
|
-
from your code, so put the policy there.
|
|
326
|
+
Promises: `Plugmem.open`, `remember`, `rememberMany`, `revise`, `recall`,
|
|
327
|
+
`forget`, `link`, `unlink`, `export`, `exportPage`, `verify`, `maintain`,
|
|
328
|
+
`checkpoint`, and every `Workspace` verb except `closeIdle`, `openCount` and
|
|
329
|
+
`close`.
|
|
216
330
|
|
|
217
|
-
|
|
218
|
-
|
|
219
|
-
|
|
220
|
-
opened. `closeIdle()` and `openCount()` stay synchronous.
|
|
331
|
+
Synchronous: `path`, `get`, `stats`, `tagsOf`, `generation`, `refresh`, `close`.
|
|
332
|
+
These touch mapped memory and return in microseconds, where a promise would be
|
|
333
|
+
pure ceremony.
|
|
221
334
|
|
|
222
|
-
|
|
223
|
-
|
|
335
|
+
Arguments are still checked on your thread: a refused one **throws** at the call
|
|
336
|
+
site rather than rejecting later, so a mistake in your code and a failure in the
|
|
337
|
+
engine never arrive the same way.
|
|
224
338
|
|
|
225
|
-
|
|
339
|
+
### Two costs a promise does not hide
|
|
226
340
|
|
|
227
|
-
|
|
228
|
-
|
|
229
|
-
loop
|
|
230
|
-
|
|
341
|
+
**The worker pool is shared, and it has four threads by default.** libuv runs
|
|
342
|
+
`fs`, `dns.lookup`, `zlib` and `crypto.pbkdf2` on the same pool this addon uses.
|
|
343
|
+
The event loop stays free either way, but four concurrent plugmem tasks fill the
|
|
344
|
+
default pool and everything else queues behind them. Measured on one machine, a
|
|
345
|
+
4 MiB `fs.readFile` in the same process:
|
|
231
346
|
|
|
232
|
-
|
|
233
|
-
|
|
347
|
+
| | `fs.readFile` |
|
|
348
|
+
|---|---|
|
|
349
|
+
| idle pool | 1.7 ms |
|
|
350
|
+
| 4 plugmem tasks in flight | 29 577 ms |
|
|
351
|
+
| the same, `UV_THREADPOOL_SIZE=8` | 1.4 ms |
|
|
352
|
+
|
|
353
|
+
plugmem's tasks are unusually long — `maintain('full')` is minutes on a large
|
|
354
|
+
memory — so raise `UV_THREADPOOL_SIZE` if the process does anything else with
|
|
355
|
+
libuv while maintenance runs.
|
|
356
|
+
|
|
357
|
+
The pool is also the ceiling on **concurrent embedding**. With an `[embedder]`
|
|
358
|
+
configured, each `remember`/`recall` occupies one worker for its HTTP round
|
|
359
|
+
trip, so at the default four, four is as parallel as it gets. Against a mock
|
|
360
|
+
provider with a fixed 100 ms latency, 16 concurrent recalls took 404 ms on the
|
|
361
|
+
default pool and 101 ms at `UV_THREADPOOL_SIZE=16` — the same 16 requests, four
|
|
362
|
+
waves or one. If a process issues many concurrent recalls against a remote
|
|
363
|
+
provider, size the pool for that, not for the CPU.
|
|
364
|
+
|
|
365
|
+
**`export()` builds its whole result on your thread.** The scan is on a worker,
|
|
366
|
+
but every fact becomes a JavaScript object during the promise's resolution, and
|
|
367
|
+
that part is main-thread work by definition. On 100 000 facts it holds the
|
|
368
|
+
thread for about 244 ms of the call's 289 ms. `exportPage()` over the same
|
|
369
|
+
memory holds it for **0 ms**, in 128-fact pages:
|
|
234
370
|
|
|
235
371
|
```ts
|
|
236
372
|
let cursor: number | undefined;
|
|
237
373
|
do {
|
|
238
374
|
const page = await db.exportPage(cursor);
|
|
239
|
-
for (const fact of page.facts)
|
|
240
|
-
await destination.write(fact);
|
|
241
|
-
}
|
|
375
|
+
for (const fact of page.facts) await destination.write(fact);
|
|
242
376
|
cursor = page.nextCursor;
|
|
243
377
|
} while (cursor !== undefined);
|
|
244
378
|
```
|
|
245
379
|
|
|
246
|
-
|
|
247
|
-
|
|
248
|
-
|
|
249
|
-
|
|
380
|
+
Each promise owns exactly one page and resolves only after its native scan
|
|
381
|
+
completed; no database lock is held while your loop body runs. A writer may
|
|
382
|
+
change between pages, so do not mutate it during a snapshot-style dump — a
|
|
383
|
+
read-only handle pages one immutable checkpoint and is stable.
|
|
250
384
|
|
|
251
|
-
|
|
252
|
-
then resolves with outcomes in input order. A maintenance call may also return
|
|
253
|
-
a no-op report when there is nothing to purge, reindex or optimize.
|
|
385
|
+
### Concurrency
|
|
254
386
|
|
|
255
|
-
`
|
|
256
|
-
|
|
257
|
-
|
|
258
|
-
|
|
387
|
+
A `Plugmem` handle is safe to use from anywhere in your process. Reads run
|
|
388
|
+
concurrently; writes serialize behind the engine's lock for the microseconds
|
|
389
|
+
they take. A second *process* opening the same file for writing is refused with
|
|
390
|
+
`PLUGMEM_LOCKED` rather than corrupting it, while any number of read-only
|
|
391
|
+
handles map the same file at once — a writer and its readers coexist across
|
|
392
|
+
processes, sharing the OS page cache.
|
|
259
393
|
|
|
260
|
-
|
|
394
|
+
## Many memories in one directory
|
|
261
395
|
|
|
262
|
-
|
|
263
|
-
|
|
264
|
-
`forget`, `link`, `unlink`, `verify`, `export`, `path`, `generation`,
|
|
265
|
-
`refresh` — is synchronous.
|
|
396
|
+
**Default: one memory, one file.** `Plugmem.open(path)` and nothing here
|
|
397
|
+
applies.
|
|
266
398
|
|
|
267
|
-
The
|
|
268
|
-
|
|
269
|
-
|
|
270
|
-
|
|
271
|
-
`remember` and a text `recall` each cost a request to the provider — so they run
|
|
272
|
-
on a libuv worker and hand JavaScript a promise. The rest touch only mapped
|
|
273
|
-
memory and return in microseconds, where a promise would be pure ceremony.
|
|
399
|
+
The problem this solves: a process serving many conversations, tenants or
|
|
400
|
+
projects wants each to have its **own** memory — nothing from one surfacing in
|
|
401
|
+
another — without managing a pile of file paths by hand. Give a name, get a
|
|
402
|
+
memory:
|
|
274
403
|
|
|
275
|
-
|
|
276
|
-
|
|
277
|
-
engine never arrive the same way.
|
|
404
|
+
```ts
|
|
405
|
+
import { Workspace, type DbEntry } from "plugmem";
|
|
278
406
|
|
|
279
|
-
|
|
407
|
+
const ws = new Workspace("/srv/memories");
|
|
280
408
|
|
|
281
|
-
|
|
282
|
-
|
|
409
|
+
// The same `Plugmem` class comes back, so a named memory has exactly the verbs
|
|
410
|
+
// a path-opened one has. A first write to an unused name creates it.
|
|
411
|
+
const chat = await ws.open("chat-42");
|
|
412
|
+
await chat.remember({ text: "prefers tokio" });
|
|
283
413
|
|
|
284
|
-
|
|
285
|
-
const
|
|
286
|
-
await
|
|
287
|
-
|
|
414
|
+
// Another name is another memory. They cannot see each other.
|
|
415
|
+
const other = await ws.open("chat-99");
|
|
416
|
+
(await other.recall({ query: "tokio" })).facts.length; // 0
|
|
417
|
+
```
|
|
418
|
+
|
|
419
|
+
Memories are **independent by design**: nothing searches across them and no
|
|
420
|
+
entity links between them. A fact filed under the wrong name is not merely
|
|
421
|
+
misplaced, it is unreachable from the other memory.
|
|
422
|
+
|
|
423
|
+
If you do not know the name, ask what each memory is for. Descriptions are
|
|
424
|
+
searchable, and so are owners, even though an owner is stored as a graph edge
|
|
425
|
+
rather than as text:
|
|
426
|
+
|
|
427
|
+
```ts
|
|
428
|
+
await ws.describe("chat-42", { description: "release planning", owner: "ann" });
|
|
429
|
+
|
|
430
|
+
const hits: DbEntry[] = await ws.find("release planning"); // → [{ db: "chat-42", … }]
|
|
431
|
+
const byOwner: DbEntry[] = await ws.find("ann"); // → the same memory
|
|
288
432
|
```
|
|
289
433
|
|
|
290
|
-
|
|
291
|
-
|
|
292
|
-
|
|
293
|
-
|
|
434
|
+
A name is `[a-z0-9][a-z0-9_-]*` and **cannot express a path**, so it resolves to
|
|
435
|
+
exactly one file inside the directory — traversal is not filtered out, it is
|
|
436
|
+
unconstructible. `ws.open(name, false)` refuses a name that does not exist yet,
|
|
437
|
+
which is what a read should do so a typo is diagnosed rather than answered with
|
|
438
|
+
an empty result.
|
|
439
|
+
|
|
440
|
+
**Who may reach which memory is not this package's job.** The name comes from
|
|
441
|
+
your code, so the policy belongs there.
|
|
442
|
+
|
|
443
|
+
### Running it
|
|
444
|
+
|
|
445
|
+
| Method | Does |
|
|
446
|
+
|---|---|
|
|
447
|
+
| `open(name, create?)` | open (default: create if missing) and hand back a `Plugmem` |
|
|
448
|
+
| `list()` | every memory in the directory, from the filesystem — including undescribed ones |
|
|
449
|
+
| `entries()` | every described memory, from the registry |
|
|
450
|
+
| `find(query, k?)` | memories whose description or owner best matches |
|
|
451
|
+
| `describe(name, args)` | record what a memory is for; revises rather than duplicating |
|
|
452
|
+
| `archive(name)` | label it archived, keeping its description. Nothing is moved or deleted |
|
|
453
|
+
| `reindex()` | rebuild the registry from the memories' own descriptions |
|
|
454
|
+
| `verify()` | report disagreements between registry and directory; repairs nothing |
|
|
455
|
+
| `closeIdle()` | close memories unused past the idle timeout (sync) |
|
|
456
|
+
| `openCount()` | how many are open right now (sync) |
|
|
457
|
+
| `close()` | close every pooled memory and the registry |
|
|
458
|
+
|
|
459
|
+
`closeIdle()` matters more than it looks. An open memory holds its file's
|
|
460
|
+
exclusive lock, so a long-running process that never lets go makes its memories
|
|
461
|
+
unreachable from anything else on the machine. Call it on a timer — that is what
|
|
462
|
+
the idle timeout is for, liveness rather than memory. The pool bounds how many
|
|
463
|
+
stay open at once (`maxOpen`, default 16, least-recently-used closed to make
|
|
464
|
+
room):
|
|
465
|
+
|
|
466
|
+
```ts
|
|
467
|
+
const ws = new Workspace("/srv/memories", { maxOpen: 16, idleTimeoutMs: 60_000 });
|
|
468
|
+
setInterval(() => ws.closeIdle(), 30_000);
|
|
469
|
+
```
|
|
470
|
+
|
|
471
|
+
A `Plugmem` handed out by `open()` is **not** closed by `ws.close()`: it is its
|
|
472
|
+
own handle holding its own lock until you close it or it is garbage collected.
|
|
473
|
+
|
|
474
|
+
`verify()` reports and never repairs, because a workspace is a directory a
|
|
475
|
+
person can edit, and guessing at their intent is how a consistency check loses
|
|
476
|
+
data.
|
|
477
|
+
|
|
478
|
+
## What it is not for
|
|
479
|
+
|
|
480
|
+
plugmem is for a local agent's memory: one process, one file, no service to
|
|
481
|
+
operate. Its design centre is around 100 000 active facts on one machine, and
|
|
482
|
+
the benchmarks track 1M-operation profiles to show how the same engine behaves
|
|
483
|
+
under heavier local load.
|
|
484
|
+
|
|
485
|
+
It is **not** a vector database and not built for multi-million vector
|
|
486
|
+
workloads, cluster sharding, multi-tenant serving or managed nearest-neighbour
|
|
487
|
+
search. For those, use a dedicated system — [Qdrant](https://qdrant.tech),
|
|
488
|
+
[Milvus](https://milvus.io), [Weaviate](https://weaviate.io),
|
|
489
|
+
[Pinecone](https://www.pinecone.io) or
|
|
490
|
+
[pgvector](https://github.com/pgvector/pgvector).
|
|
491
|
+
|
|
492
|
+
## Other ways in
|
|
493
|
+
|
|
494
|
+
The same engine ships four ways. This package is the Node one.
|
|
495
|
+
|
|
496
|
+
| You are | Use |
|
|
497
|
+
|---|---|
|
|
498
|
+
| writing JavaScript / TypeScript for Node | **this package** |
|
|
499
|
+
| writing Rust | [`plugmem-host`](https://docs.rs/plugmem-host/latest) — the engine in your process |
|
|
500
|
+
| an agent, or another language | [`plugmem-mcp`](https://docs.rs/plugmem-mcp/latest) — a stdio JSON-RPC sidecar |
|
|
501
|
+
| a person at a terminal | [`plugmem-cli`](https://docs.rs/plugmem-cli/latest) |
|
|
294
502
|
|
|
295
|
-
|
|
296
|
-
|
|
297
|
-
|
|
503
|
+
**Working with an LLM agent?** There is a companion
|
|
504
|
+
[skill](https://github.com/m62624/plugmem/blob/main/skill/SKILL.md) describing
|
|
505
|
+
the remember/recall loop, the contradiction workflow and the verbs. This package
|
|
506
|
+
ships it: `skill()` returns the text and `skillVersion()` the version it was
|
|
507
|
+
written against.
|
|
298
508
|
|
|
299
509
|
## License
|
|
300
510
|
|
|
301
|
-
MIT.
|
|
511
|
+
MIT. Source: <https://github.com/m62624/plugmem>
|
package/index.d.ts
CHANGED
|
@@ -86,6 +86,18 @@ export interface RecallArgs {
|
|
|
86
86
|
k?: number
|
|
87
87
|
/** Include closed revisions (default false). */
|
|
88
88
|
closed?: boolean
|
|
89
|
+
/**
|
|
90
|
+
* Token budget of the `rendered` block (default 512). That block is what
|
|
91
|
+
* goes into a prompt, so this is the knob deciding how much of the
|
|
92
|
+
* context window a recall may spend.
|
|
93
|
+
*/
|
|
94
|
+
tokenBudget?: number
|
|
95
|
+
/**
|
|
96
|
+
* HNSW beam width for the vector source (default: the configured
|
|
97
|
+
* `hnsw_ef_search`). Higher is more accurate and slower; ignored while
|
|
98
|
+
* the engine is still in the flat regime, below `flat_to_hnsw`.
|
|
99
|
+
*/
|
|
100
|
+
ef?: number
|
|
89
101
|
/**
|
|
90
102
|
* A precomputed embedding. Its length must equal the configured `dim`.
|
|
91
103
|
*
|
|
@@ -104,6 +116,12 @@ export interface LinkArgs {
|
|
|
104
116
|
rel: string
|
|
105
117
|
/** Destination entity name. */
|
|
106
118
|
dst: string
|
|
119
|
+
/**
|
|
120
|
+
* The fact this edge follows from, recorded on the edge and returned by
|
|
121
|
+
* graph recall — the answer to "why is this edge here". Ignored by
|
|
122
|
+
* `unlink`, which closes an edge rather than opening one.
|
|
123
|
+
*/
|
|
124
|
+
provenance?: number
|
|
107
125
|
}
|
|
108
126
|
/** One similar / potentially-conflicting live fact surfaced by `remember`. */
|
|
109
127
|
export interface Similar {
|
|
@@ -226,6 +244,12 @@ export interface FactSnapshot {
|
|
|
226
244
|
}
|
|
227
245
|
/** One exported fact — the id-free, import-ready shape. */
|
|
228
246
|
export interface ExportedFact {
|
|
247
|
+
/**
|
|
248
|
+
* The fact's id in the database it came from. Informational: an import
|
|
249
|
+
* assigns fresh ids. Present because edges name their provenance fact by
|
|
250
|
+
* id, so a dump carrying edges needs something for them to point at.
|
|
251
|
+
*/
|
|
252
|
+
id: number
|
|
229
253
|
/** The fact text. */
|
|
230
254
|
text: string
|
|
231
255
|
/** Subject entity name, if any. */
|
|
@@ -549,10 +573,14 @@ export declare class Plugmem {
|
|
|
549
573
|
/**
|
|
550
574
|
* Every currently-open fact, as one array (id-free, import-ready).
|
|
551
575
|
*
|
|
552
|
-
* **Async, but still unbounded
|
|
553
|
-
* whole memory is materialized into a single array before it
|
|
554
|
-
*
|
|
555
|
-
*
|
|
576
|
+
* **Async, but still unbounded, in two ways.** The scan runs on a worker,
|
|
577
|
+
* yet the whole memory is materialized into a single array before it
|
|
578
|
+
* resolves, so peak memory is the whole export. And `resolve` runs on the
|
|
579
|
+
* JS thread by definition, so building one JavaScript object per fact
|
|
580
|
+
* stalls it: measured at ~244 ms of a 289 ms call over 100 000 facts,
|
|
581
|
+
* against 0 ms for the same data through `exportPage`. A promise hides
|
|
582
|
+
* neither cost. Prefer `exportPage` for anything but a small memory or a
|
|
583
|
+
* script.
|
|
556
584
|
*/
|
|
557
585
|
export(): Promise<ExportedFact[]>
|
|
558
586
|
/**
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "plugmem",
|
|
3
|
-
"version": "0.
|
|
3
|
+
"version": "0.5.0",
|
|
4
4
|
"description": "Native Node.js addon for plugmem: an embedded long-term memory engine for LLM agents (remember / recall / revise / forget over one local file).",
|
|
5
5
|
"repository": {
|
|
6
6
|
"type": "git",
|
|
@@ -41,11 +41,11 @@
|
|
|
41
41
|
"typecheck": "tsc -p tsconfig.json"
|
|
42
42
|
},
|
|
43
43
|
"optionalDependencies": {
|
|
44
|
-
"plugmem-linux-x64-gnu": "0.
|
|
45
|
-
"plugmem-linux-arm64-gnu": "0.
|
|
46
|
-
"plugmem-darwin-x64": "0.
|
|
47
|
-
"plugmem-darwin-arm64": "0.
|
|
48
|
-
"plugmem-win32-x64-msvc": "0.
|
|
49
|
-
"plugmem-win32-arm64-msvc": "0.
|
|
44
|
+
"plugmem-linux-x64-gnu": "0.5.0",
|
|
45
|
+
"plugmem-linux-arm64-gnu": "0.5.0",
|
|
46
|
+
"plugmem-darwin-x64": "0.5.0",
|
|
47
|
+
"plugmem-darwin-arm64": "0.5.0",
|
|
48
|
+
"plugmem-win32-x64-msvc": "0.5.0",
|
|
49
|
+
"plugmem-win32-arm64-msvc": "0.5.0"
|
|
50
50
|
}
|
|
51
51
|
}
|