[!NOTE] 📦 This plugin lives in the dsh-plugins monorepo — file issues & pull requests there. npm: @hy-sde-org/dsh-memory-extraction

dsh-memory-extraction — automatic long-term-memory extraction for DeepSeek Harness

Standalone package @hy-sde-org/dsh-memory-extraction: automatic long-term-memory extraction at compaction checkpoints for DeepSeek Harness. After every compaction/summary boundary a bounded, fail-open pipeline projects the user-authored text of the checkpointed range, proposes and canonicalizes durable facts with auxiliary model calls (evidence-grounding, proposal/admission/canonicalization, per-session cursors, cross-session evidence floor), and commits admitted facts into the same project memory bank retain/learn write — strictly additive to the explicit memory surface (retain / recall / reflect / memory_edit / learn stay the model-facing path).

Identity	Value
Package	@hy-sde-org/dsh-memory-extraction
Plugin id	memory-extraction (the row you add by hand — the package ships no bundle row)
Plane	host — injects memory + llm, publishes nothing
Commit seam	ctx.memory (extracted facts land with source: 'memory_extract')

Based on Apache Maka (Apache License 2.0, © the Apache Software Foundation) — the memory-extraction engine (evidence projection, proposal/admission/canonicalization, control store, compaction trigger) is ported from Maka (No. 2), upstream packages/runtime/src/memory-extraction.ts. The DSH integration is derived from the DeepSeek Harness codebase (MIT, © 2026 DeepSeek); each derived file carries the attribution in its header. See THIRD-PARTY-NOTICES.md.

Why

Without this package, durable memory only grows when the model remembers to call retain/learn. This engine watches every compaction checkpoint instead: facts the user states across a session are projected, verified, and committed automatically, so memory survives compaction without manual discipline. The load-bearing rules are ported from Maka and enforced by the engine:

• Evidence is user-authored text only — assistant text is interpretation-only; tool calls/results and reasoning are opaque. Evidence is bounded (12 000 chars of JSON / 4 000 chars per record) and fail-closed on overflow.
• Admission is verified — proposal → admission (verbatim-quote check against the bounded evidence + secret rejection) → canonicalization → re-admission, at most 3 auxiliary model calls per range with a 60 s timeout; failures are contained (fail-open at the runtime boundary — compaction is never blocked).
• The cursor only moves to a committed boundary — write order is items → cursor → receipt, so a crash can never double-process; operation ids are deterministic (memory_ + sha256 of {sessionId, trigger, boundarySeq}), and receipts make replays no-ops.
• Subagents are excluded — child-agent compactions never extract (excludeSubagents: true default).
• Cross-session evidence floor — a fact only commits once minGapEvidence distinct sessions (default 2) have proposed the same durable fact; uncorroborated facts stay in the gap ledger as open pending facts and expire after gapLedgerMaxAgeMs (default 90 d).

Prerequisites

• Node.js 22.19 or newer with npm and pnpm on PATH;
• a DeepSeek Harness release carrying the 0.2.0-rc.2 peer range — @deepseek-ai/cordis ~4.0.4 and @deepseek-ai/dsh-compaction, @deepseek-ai/dsh-llm, @deepseek-ai/dsh-session, @deepseek-ai/dsh-storage ^0.2.0-rc.2 — including the standard dsh CLI;
• the @hy-sde-org/dsh-memory plugin mounted — extraction commits into its ctx.memory bank;
• a storage backend exposing a kv facet — both shipped backends (sqlite and storage-json) do. Nothing else: no API keys (the auxiliary calls reuse the session's routed model unless you configure provider/model).

Install the Harness CLI and pnpm before continuing:

npm install --global @deepseek-ai/dsh pnpm
dsh --version

Quick start

The package ships no dsh.bundle.patch, so mounting is two steps: install the package, then add one row to your profile's host composition by hand.

Route A — published npm package

dsh plugin --profile web add @hy-sde-org/dsh-memory-extraction

Because the package declares no bundle, the CLI installs it as a plain profile dependency and warns declares no dsh.bundle — installed as a plain dependency, not a profile layer. No row is applied automatically — add it to the host composition (the profile's cordis.patch.yml patch layer, loaded before any session):

- insert:
    - id: memory-extraction
      name: '@hy-sde-org/dsh-memory-extraction'
      config:
        backend: sqlite           # storage backend hosting the control unit (default sqlite)
        enabled: true             # master switch (default true)
        excludeSubagents: true    # child-agent compactions never extract (default true)
        # provider: <cheap provider id>  # optional; unset uses the session's routed model config
        # model: <cheap model id>        # optional; unset uses the session's routed model config
        importance: 0.5           # bank importance for auto-extracted facts (default 0.5)
        dedupe: true              # probe the bank before committing duplicates (default true)
        timeoutMs: 60000          # auxiliary call timeout (default 60000)
        minGapEvidence: 2         # distinct sessions before a fact commits (default 2; 0 disables the floor)
        gapLedgerMaxAgeMs: 7776000000   # sightings older than this stop counting (default 90 d in ms)

Route B — from source (validate this checkout or hack on the plugin)

git clone git@github.com:hy-sde/dsh-plugins.git
cd dsh-plugins
pnpm install
PACKAGE_TARBALL="$(cd dsh-memory-extraction/packages/memory-extraction && pnpm pack --pack-destination /tmp | tail -n 1)"
dsh plugin --profile web add "$PACKAGE_TARBALL"

pnpm pack runs the normal prepack build and produces a tarball containing dist/. The install lands as a plain profile dependency exactly as in Route A — add the host-composition row above the same way.

Verify the composed configuration

dsh web --dump-config

The composed tree must show the memory-extraction row loading @hy-sde-org/dsh-memory-extraction.

Run

dsh web

After the next compaction checkpoint that contains user-authored text, the engine proposes, verifies, and commits durable facts into the project memory bank with source: 'memory_extract' — ask the agent to recall or reflect and the auto-extracted entries are there. The plugin boots asynchronously; if the configured storage backend is missing it logs and stays inert instead of failing composition.

Uninstall

Remove the row you added to the host composition, then:

dsh plugin --profile web remove @hy-sde-org/dsh-memory-extraction

What the package does

The engine is a host-plane Cordis plugin (inject: ['memory', 'llm'], it publishes nothing): one process opens the memory_extraction control unit once (tables cursors / receipts / failures / gaps over a KvUnit) and observes every session's events through an unscoped ctx.on('session/event') listener. On each compaction/summary boundary it runs the per-session serialized pipeline — project evidence, propose, admit, canonicalize, re-admit, commit — and never blocks the compaction listener. Committed facts go through the ctx.memory commit surface with a dedupe probe, landing in the same bank retain/learn write; deferred facts stay sighted in the gap ledger and are shown to later proposal passes (the model cites a gapId instead of coining a paraphrase). The engine itself is pure and testable without cordis:

import { MemoryExtractionEngine } from '@hy-sde-org/dsh-memory-extraction'

const engine = new MemoryExtractionEngine(ports) // readGate/readEvents/cursor+receipt+failure/commitItems/generate
const result = await engine.execute(snapshot)   // never throws; idempotent by operation id

Configuration

All options are optional; the row's config: fills the defaults.

Option	Default	Purpose
backend	sqlite	storage backend hosting the control unit; must expose a kv facet
enabled	true	master switch
excludeSubagents	true	child-agent compactions never extract
provider, model	unset	auxiliary model route; unset uses the session's routed request header. A cheap auxiliary model is recommended for large deployments
importance	0.5	bank importance for auto-extracted facts
dedupe	true	probe the bank before committing duplicates
timeoutMs	60000	per auxiliary-call timeout; a range makes up to 3 auxiliary calls
minGapEvidence	2	distinct sessions before a fact commits; 0 disables the floor
gapLedgerMaxAgeMs	7776000000	sightings older than this stop counting (90 d); 0 disables the ledger

Compatibility

Component	Supported contract
Node.js	22.19 or newer (engines.node >=22.19.0)
DeepSeek Harness	0.2.0-rc.2 peer range (@deepseek-ai/cordis ~4.0.4; dsh-compaction, dsh-llm, dsh-session, dsh-storage ^0.2.0-rc.2)
Memory bank	@hy-sde-org/dsh-memory (ctx.memory seam)
Storage backend	must expose a kv facet — shipped sqlite and storage-json both yes

Upstream seam-contract changes require a new package release and contract review.

Development

pnpm install
pnpm check        # strict typecheck of packages/memory-extraction
pnpm test         # vitest run
pnpm build        # tsc -> dist
pnpm pack:all     # pack the publishable package

The engine is a pure state machine (src/engine.ts) over injected ports, so the state machine, evidence projection, and parsers are testable without cordis; src/runtime.ts is the only host wiring. See packages/memory-extraction/README.md (English | 中文) for the implementation tour.

License and attribution

This package is licensed MIT — see LICENSE. Code derived from the DeepSeek Harness codebase is MIT (© 2026 DeepSeek); the memory-extraction engine is a port of Apache Maka (No. 2) under the Apache License 2.0 (© the Apache Software Foundation) — the Apache License 2.0 text applies to the derived Maka modules, and the rest of the repository is MIT. The upstream notice text is reproduced in THIRD-PARTY-NOTICES.md.

This plugin is a separate installable package; the harness remains the property of its own project.