dreamweave

KO | EN

gitlite — search

#anthropic #java #shell #postgresql #list #redis #svg #pdf #vue #pytorch

(★ 9)

dreamweave - a nightly memory engine for OpenClaw-based agents (incl. Microsoft Scout). Your agent dreams: it weaves a brain-faithful graph+vector memory and consolidates it like sleep - forgetting curve, merge, temporal weave - across three tiers. Independent community project; not affiliated with or endorsed by Microsoft.

파일 탐색기

ZIP 다운로드

- - release.yml
- - adapter.mjs
- incremental-3tier-180d.json
- - memory-nebula.png
- ARCHITECTURE.md
- FIRST-PRINCIPLES.md
- JOURNEY.md
- JUDGMENT-SURFACES.md
- - install.md
- - install.md
- derived-memory.js
- dream.js
- dump-active.js
- embed.js
- entities.js
- graphtext.js
- langsvc.English.js
- langsvc.js
- memory-render.js
- recall.js
- schema.js
- sig-utils.js
- timeline.js
- tuning.js
- activation-seed.test.js
- anchor-projection.test.js
- apply-noop.test.js
- archive-recall.test.js
- chronicle-delete-cascade.test.js
- chronicle-entity-prune.test.js
- chronicle-lifecycle.test.js
- chronicle-projection-head.test.js
- chronicle-resummarize.test.js
- chronicle-retrieval-surface.test.js
- chronicle-scheduling.test.js
- derived-evidence-contract.test.js
- dream-atomicity.test.js
- embed-batch.test.js
- entities-incremental.test.js
- entity-ambiguity.test.js
- entity-apply-link.test.js
- entity-hub-review-dirty-seq.test.js
- entity-hub-review.test.js
- entity-mapping-dataflow.test.js
- entity-status-phrases.test.js
- envelope-roundtrip.test.js
- gist-evidence.test.js
- hybrid-seed.test.js
- incremental-lineage.test.js
- ingest-dates.test.js
- ingest-episodic.test.js
- langsvc-plugin.test.js
- langsvc-reembed.test.js
- merge-nondestructive.test.js
- merge-validation.test.js
- microsoft-scout-install.test.js
- parse-date-range.test.js
- processing-seq.test.js
- reactivation-order.test.js
- reactivation-synthesis.test.js
- recall-supersede.test.js
- release-manifest.test.js
- repair-dates.test.js
- run.js
- salience-score.test.js
- schema-integrity.test.js
- sequence-recall.test.js
- setup-neutral.test.js
- supersede-aspect.test.js
- temporal-chronicles.test.js
- tuning.test.js
- vector-repair.test.js
- viz-archive-exclusion.test.js
- viz-projection.test.js
- viz-template.test.js
- graph-store-visualization.html
- lib-3d-force-graph.min.js
.gitattributes
.gitignore
config.js
INSTALL.md
LICENSE
package-lock.json
package.json
plan.md
README.md
setup.js
spec-derived-memory-unification.md

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

**A host-independent nightly memory engine with adapters for Microsoft Scout and [Hermes Agent](https://hermes-agent.nousresearch.com/docs/).**

Most agent memory is a flat text file that grows unbounded and gets RAG'd. **dreamweave** is different: your agent *dreams*. A nightly pass **weaves a brain-faithful graph+vector memory** and consolidates it like sleep — a forgetting curve drops noise, duplicates **merge**, and survivors **weave** into one connected graph — across three tiers (instinct / recall / archive). The **recall** path then answers with semantic vector search **plus graph-neighbor expansion**.

<img src="docs/assets/memory-nebula.png" alt="Dreamweave semantic nebula visualizing 1,024 memories and their entity connections" width="100%">

Your memory, alive: explore facts, entities, temporal structure, and activated connections in the 3D semantic nebula.

What makes it different from other "dream"/memory skills:

- **Graph + vector, not flat RAG** — entities and facts with edges; recall returns *neighbors*, which is what actually helps synthesis. The **weave** is the signature operation.
- **Three tiers** — instinct (~500 injected gist) / recall (~2,500 graph+vector) / archive (uncapped bookshelf). It bounds what's **injected**, not what's **remembered**.
- **Brain-faithful consolidation** — forgetting curve, REM-style merge, temporal sequencing, demote-don't-delete.
- **Bounded nightly cost** — incremental weave + embed-once, so it's genuinely runnable *every* night, forever.
- **Benchmarked** — measured on the [Recall Bench](https://github.com/Stevenic/recall), not vibes.

The engine runtime is **local and free**: SQLite (`better-sqlite3`) + `sqlite-vec` for the vector index,
and `@huggingface/transformers` for on-device embeddings (`all-MiniLM-L6-v2`, 384-dim). It has no
provider client or API-key requirement; nightly judgment is supplied by the host agent. Ships with
a 3D **semantic nebula explorer** to inspect the store.

> **Not affiliated with, authorized, or endorsed by Microsoft.** This is an independent, community
> project. "Microsoft Scout" and related names are trademarks of their respective owners and are
> referenced only to describe compatibility.

---

## Why

Most agents store memory as a growing list of text snippets and retrieve with naive similarity.
That fails two ways at scale: the bank **grows unbounded** (blowing the context budget and diluting
attention), and facts sit **disconnected** (so retrieval can't hop between related facts). This
package fixes both:

- **FORGET** — a forgetting curve decays stale facts; noise evaporates, durable facts persist.
- **MERGE** — duplicate/over-lapping facts collapse into fewer, richer entries (count down, info kept).
- **WEAVE** — every surviving fact is connected into one graph via shared entities, so recall can
  traverse and attention can bind related facts.

The result is a bounded projection (~250-entry standard target) backed by a larger tiered,
atomic, **fully connected** memory store.

> **Design guideline:** the engine follows a **three-tier memory model** (instincts /
> RAG / bookshelf) with a few non-negotiable principles (embed-once, bounded nightly
> cost, demote-don't-delete, gist-for-attention/detail-for-recall). Start with
> [`docs/FIRST-PRINCIPLES.md`](docs/FIRST-PRINCIPLES.md) (why), then
> [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md) (the design) and
> [`docs/JOURNEY.md`](docs/JOURNEY.md) (how we got here) before making engine changes.

### Host judgment layer

The engine is **local-only and has no external-LLM client**. It performs embeddings, decay,
graph maintenance, candidate generation, and decision application locally. For judgments a
regex cannot make, it emits bounded `report-*` JSON; the host agent reads that report and returns
decision JSON to the matching `apply-*` command.

The caller judges six surfaces in order: entities, aliases, salience, merges, synthesis, and chronicles.
The model is a **judge, not an author**: decisions may only type, group, score, or consolidate
content present in the report. This keeps provider credentials and model choice outside the
engine and gives the live harness and evaluation harness one shared execution path.

---

## Architecture (one minute)

- **`memory.db`** (SQLite + `sqlite-vec`) is the durable source of truth and retrieval index.
- Three node kinds:
  - **fact** — an atomic memory (`kind='fact'`), carries a forgetting-curve `strength`, **projects
    back to your agent's memory bank**.
  - **entity** — a connector hub (`person:`/`team:`/`system:`/…), no decay, **never projects**;
    pure graph scaffolding that links facts so nothing is an island.
  - **chronicle** — an evidence-linked account of a day or coarser period, providing a temporal recall axis.
- A parallel `vec_nodes` (vec0, cosine, 384-dim) holds each node's embedding.
- The **host agent's memory bank is a disposable nightly projection** of the db. Daytime, the agent
  adds memories normally; the nightly **dream** ingests them, consolidates, and projects the curated
  set back.

```
  your agent's memories  ──dump──▶  snapshot.json
                                      │ ingest
                                      ▼
                                 memory.db  ──dream / weave / reflect / consolidate──▶  memory.db
                                      │ export-harness (diff)
                                      ▼
                          add/remove memories in the agent
```

---

## Install

Requirements: **Node ≥ 20** and a toolchain that can build `better-sqlite3` (prebuilt binaries cover
most platforms; otherwise you need Python + a C++ compiler).

### Option A — download a release zip (recommended for users)

Grab the latest `dreamweave-vX.Y.Z.zip` from the [**Releases**](https://github.com/spqian/dreamweave/releases)
page, unzip it, and tell your assistant to *"import this"*. The root
[`INSTALL.md`](INSTALL.md) uses the current runtime's known harness, asks only when
that identity is uncertain, and then follows the selected adapter guide.
Older release zips may predate the Hermes adapter; use a checkout or release containing
`harness-adapters/hermes-agent/`. Shared engine bootstrap can also be run by hand:

```bash
cd dreamweave-vX.Y.Z
npm install
npm run setup
```

> Release zips are built and published automatically by [`.github/workflows/release.yml`](.github/workflows/release.yml)
> whenever a `vX.Y.Z` tag is pushed. They are **source-only** — the native deps are built by your
> `npm install`. Maintainers can build the same zip locally with `npm run pack`.

### Option B — clone the repo (for development)

```bash
cd dreamweave
npm install
npm run setup
```

`npm run setup` verifies dependencies, creates the data dir, initializes a fresh `memory.db` with the
full schema, and warms the embedding model (first run downloads ~90 MB once, then it's cached).
It does not install host skills automatically. Continue with the selected adapter below.

### Harness adapters

| Target | Installation guide | Integration boundary |
| --- | --- | --- |
| Microsoft Scout | [MicrosoftScout/install.md](harness-adapters/MicrosoftScout/install.md) | Scout skill installation and native memory-tool workflow |
| Hermes Agent | [hermes-agent/install.md](harness-adapters/hermes-agent/install.md) | Profile-local memory transport, skills and checkpointed host-LLM judgment |

The Hermes adapter preserves whole `§`-delimited memory cards, uses stable content-derived
source IDs, and records a projection manifest so generated summaries are not ingested as
new observations. Its helper checks report freshness, audited decisions, engine acceptance,
sync and graph health before publishing the next projection. It **does not contain an LLM
client**: a normal Hermes agent performs the judgments. Uncertain evidence can be explicitly
deferred without mutation; bounded backlog is reported rather than called complete.

#### Hermes memory budget: characters, tokens and projected entries

There are separate limits; changing one does not change the others:

| Setting | Larger-assistant example | Meaning |
| --- | ---: | --- |
| `memory.memory_char_limit` | 250000 | Hermes `MEMORY.md` character limit, **not a token limit** |
| `memory.user_char_limit` | 15000 | Separate Hermes `USER.md` character limit |
| `MEMORY_ENTRY_TARGET` | 200 | Desired Dreamweave projection entry count |
| `MEMORY_ENTRY_MAX` | 200 | Maximum projection entry count; not total database capacity |

With user approval, configure the **selected Hermes profile**, for example:

```bash
hermes config set memory.memory_char_limit 250000
hermes config set memory.user_char_limit 15000
```

Configure the adapter's projection budget to match the intended entry target/cap, and its
publication character budget to fit the selected profile; see the adapter guide. Use the
`MEMORY_ENTRY_*` overrides consistently across a cycle, not only at export. The standard
engine capacity preset is otherwise unchanged. **200 entries is a cap/target, not a promise
that every export contains exactly 200**, nor a promise that 200 arbitrarily long entries
fit the character budget. Retained detail stays in the database for recall.

The larger profile above has been exercised with a 200-entry projection, but is **opt-in**,
not a universal default. At the rough English-text heuristic of four characters per token,
250,000 characters is about **62,500 tokens**; actual tokenization depends on the model,
language and content. Leave room for the conversation, tools, reports, reasoning and output.
Larger injected memory increases latency, cost and compaction pressure. For a small-context
model, select a smaller projection and rely more on graph recall. Do not disable compaction
or silently truncate memory to fit. Judgment and context-compression models are separate
Hermes settings; verify both are available through the chosen provider.

### Where data lives (all overridable)

| Env var | Default | What |
| --- | --- | --- |
| `DREAM_MEMORY_DIR` (or `AGENT_MEMORY_DIR`) | `~/.dream-memory` | per-user data dir (db, model cache, rendered viz) |
| `MEMORY_DB` | `<dir>/memory.db` | the SQLite store |
| `MEMORY_VIZ` | `<dir>/memory-graph.html` | rendered 3D explorer output |
| `MEMORY_MODEL_CACHE` | `<dir>/model-cache` | embedding model weights |
| `MEMORY_MODEL` | `Xenova/all-MiniLM-L6-v2` | embedding model id |
| `MEMORY_EMBED_DIM` | `384` | embedding dimensionality (must match the model) |

To co-locate with a host agent's data dir, set `AGENT_MEMORY_DIR` to it (e.g. `~/.copilot/data`).

---

## Usage

### Recall (read path)

```bash
node src/recall.js --query "who owns the gateway release" --max-hops 2
```

Returns semantic seeds (cosine similarity) plus a connected cluster (nodes + typed edges) for the
agent to ground an answer on. Drives the **`graph-recall`** skill.

### The nightly dream (maintenance path)

Run these in order (the **`dream`** skill orchestrates them with the host's memory tools):

```bash
# 1. WAKE: dump your agent's memories to snapshot.json  →  [{ id, fact, category }]
#    category ∈ decision | fact | context | preference  (a DISPLAY label only)
#    NB: category no longer sets the memory class — every ingested memory enters EPISODIC;
#    the nightly dream EARNS semantic durability through reactivation and assigns
#    continuous salience_score importance only through the salience judgment surface.

# 2. INGEST + verify (lossless, memory_id-keyed)
node src/dream.js ingest-harness --file snapshot.json
node src/dream.js verify-sync   --file snapshot.json   # exit 3 if any memory is missing

# 3. DREAM: pre-weave, decay, reactivate, evaporate, housekeeping
node src/dream.js dream

# 4. REPORT → HOST JUDGES → APPLY, in this order:
# entities, aliases, salience, merges, synthesis, chronicles
node src/dream.js report-entities --as-of <iso>
node src/dream.js apply-entities --file decisions.json --as-of <iso>
# ...repeat for aliases, salience, merges, synthesis, chronicles
# For merges, preserve the report identity:
# report-merges -> { report_id, clusters }
# apply-merges file -> { report_id, decisions:[...] }
# Merge apply is all-or-nothing and exits 3 with structured rejections if
# any submitted decision is stale, overlapping, or spans report clusters.

# 5. HEALTH: connect post-judgment changes and verify invariants
node src/dream.js weave --as-of <iso>
node src/dream.js doctor # exit 3 on islands/dangling/vector defects

# 6. PROJECT: export the curated facts back to the agent's bank (apply the diff)
node src/dream.js export-harness

# 7. VIZ
node src/dream.js export-viz          # renders $MEMORY_VIZ
```

Helpers: `node src/dream.js stats` · `budget` (entry-count pressure + forecast) · `init` (just create
the db).

See **`docs/JUDGMENT-SURFACES.md`** for the host-neutral report/judgment/apply contract.
Runnable `dream` and `graph-recall` skills belong to the selected host adapter; the repository
root deliberately exports no globally named skills.

---

## The 3D explorer

`export-viz` renders a self-contained HTML explorer (`$MEMORY_VIZ`) plus its vendored engine next to
it. Open the file in any browser. Features: a hub-anchored graph layout with a linear semantic
projection fallback, color = episodic/semantic durability plus continuous salience highlighting,
white balls for entity hubs, deep search over title+content,
click-to-focus with 2-hop pruning and a 3D "auto-dodge" that pushes irrelevant nodes out of view,
minimap, orbit controls, and smart auto-labeling. Append `?scoutTheme=dark` or `?scoutTheme=light`
to force a theme.

---

## Integrating with a host agent

The engine is **host-agnostic** — it only reads a snapshot JSON and writes an inject-ready export.
To wire it into an agent:

1. **Select a host adapter.** Install its own `dream` and `graph-recall` skills. Do not install
   another adapter's same-named skills into the same namespace. The adapter binds the
   host-neutral engine contract to that harness's transport, projection, and tool model.
2. **Follow that adapter's guide.** It owns memory transport, projection, skill installation,
   scheduling, and host-specific safety gates. For a new harness, implement these against the
   engine's snapshot/export interface and the canonical judgment contract.
3. **Co-locate data** via `AGENT_MEMORY_DIR` if you want the store beside the agent's other state.

---

## Layout

```
agent-memory/
  config.js                 # env-overridable paths + model config
  setup.js                  # host-neutral engine bootstrap
  package.json
  INSTALL.md                # selects the known harness; asks only if uncertain
  harness-adapters/
    MicrosoftScout/         # Scout installation guide, installer, dream + recall skills
    hermes-agent/           # Hermes installer, transport, dream + recall skills, helper + tests
  src/
    dream.js                # the consolidation engine (all subcommands)
    recall.js               # vector + graph recall (read path) — temporal parsing/tokenization via langsvc
    embed.js                # local embeddings (transformers.js)
    entities.js             # backward-compat shim (re-exports sig-utils + default langsvc)
    sig-utils.js             # generic "type:slug" signature parsing (labelOf/typeOf/buildVocab only)
    langsvc.js               # pluggable language-service loader/facade (no LLM, local, deterministic)
    langsvc.English.js       # default (English) service: entities, normalize/slug, temporal parsing,
                             #   tokenization/stopwords, query-shape detection, hard-specifics, age-tag, node prose
    timeline.js              # age-day math (generic) + ageTag/relAge facade (labels resolved via langsvc)
    graphtext.js            # neighbor-naming text for embeddings — facade over langsvc's renderNodeText
    schema.js               # fresh-db schema bootstrap
  viz/
    graph-store-visualization.html   # explorer template (empty data line)
    lib-3d-force-graph.min.js        # vendored 3D engine (MIT)
  docs/
    JUDGMENT-SURFACES.md    # canonical host-neutral report/judge/apply contract
```

---

## Licenses

MIT (this package). Bundled/declared dependencies: `3d-force-graph` (MIT), `better-sqlite3` (MIT),
`sqlite-vec` (MIT/Apache-2.0), `@huggingface/transformers` (Apache-2.0), and the
`all-MiniLM-L6-v2` model weights (Apache-2.0). Verify before redistribution.

dreamweave

파일 탐색기

# CDN으로 사용하기

Was this content helpful?

Recent Feedback

Download README