KLRSL/dsh-biomemory

中文文档 · English

A cross-session memory plugin for DeepSeek Harness (DSH), designed like a human brain: layered memory, graded approval, memory metabolism, fully transparent and editable.

• Plain Markdown data layer (default ~/.dsh/memory, overridable via the DSH_MEMORY_ROOT env var) — human-readable, edit-and-take-effect
• memory tool: add / query / remove / list / pin / unpin / dream / audit
• Frozen snapshot injection at session start (pinned memories and user preferences at top priority, then recent knowledge/behavior)
• Graded approval gate: important memories (preferences/decisions/lessons) require human approval; ordinary facts are auto-saved; fails closed when no approval channel is available
• Audit: human-readable audit.log (legacy, compatible) + structured audit.jsonl (JSON Lines) — every event traceable
• /memory command: list / query / add / remove / pin / unpin / dream / audit
• memory_recall tool: cross-session recall ("do you remember…" scenarios)
• Deduplication: content fingerprint skips duplicate entries
• Memory metabolism (/memory dream): half-life decay, reference consolidation, conflict arbitration, low-weight archiving — with automatic backup & rollback
• Memory pins: lock a memory so it never decays and always enters the snapshot
• Semantic retrieval: pure-JS TF-IDF + cosine — no native modules, no external dependencies

Install

# As a local bundle in a DSH profile
dsh plugin add dsh-biomemory
# Or pnpm local link
pnpm add link:./dsh-biomemory

Add dsh-biomemory to dsh.profile.bundles in the profile.

Memory Layout

~/.dsh/memory/
├── preferences.md      # User/project preferences (top priority, frozen-injected)
├── hot/
│   ├── knowledge.md    # L1 recent knowledge (facts/decisions)
│   └── behavior.md     # L1 recent behavior (lessons/habits/workflows)
├── projects/<name>/    # L2 project archives
├── longterm/           # L3 long-term memory
├── archive/            # Memories archived by metabolism (decayed below threshold, never deleted)
├── backups/            # Automatic backups before dream runs (rollback source)
├── audit.log           # Human-readable audit (legacy, kept for compatibility)
└── audit.jsonl         # Structured audit (JSON Lines, v0.3)

Each entry is a single line: - [knowledge|auto] [fp:xxx] [w:10] [h:3] [t:2026-08-16 13:00] [pin] text

• w = weight (default 10) — decay/consolidation base
• h = reference count — consolidation input
• t = write time — decay age source
• pin = locked (excluded from decay, always injected)

Memory Metabolism (Dream)

/memory dream (or memory action=dream) manually triggers memory metabolism — the housekeeping a sleeping brain does:

1. Half-life decay (default 7 days): weight halves every half-life (w × 0.5^(age/halfLife)), floored at 1.
2. Reference consolidation: entries referenced ≥ consolidateThreshold (default 3) times gain +1 weight, capped at weightCap (default 20).
3. Conflict arbitration: when behavior memory conflicts with preferences, preferences win — the behavior entry's weight is halved and a CONFLICT audit event is recorded.
4. Archiving: entries whose weight drops below decayThreshold (default 3) move to archive/ — moved, never deleted.

Usage:

/memory dream            # run metabolism
/memory dream --dry-run  # preview only, no changes
memory action=dream dryRun=true   # same via the memory tool

Dry-run example output:

【预览】扫描 120 条：衰减 12 · 巩固 3 · 冲突 0 · 归档 4
备份：（dry-run 不执行备份）

Backup & rollback: before an actual run, the whole memory store is automatically copied to backups/<timestamp>/ (including audit.jsonl). On startup, the self-check restores the latest backup automatically if a primary memory file is found corrupted. Rollbacks are recorded as ROLLBACK audit events.

Memory Pins

Lock a memory so it never participates in decay and always enters the snapshot:

/memory pin <fp>      # lock
/memory unpin <fp>    # unlock
memory action=pin fp="xxx"
memory action=unpin fp="xxx"

Snapshot injection priority: pinned > preferences > knowledge > behavior.

Audit

Two audit channels:

• audit.log — human-readable one-line summaries, backward compatible
• audit.jsonl — structured, one JSON object per line

Events: WRITE, DECAY, CONSOLIDATE, CONFLICT, ARCHIVE, PIN, UNPIN, PREVIEW (dry-run), ROLLBACK.

Example line:

{"t":"2026-08-16T05:00:00.000Z","event":"DECAY","fp":"abc123","text":"..."}

Query:

/memory audit                    # recent events
/memory audit --since 7d         # last 7 days
/memory audit --type DECAY       # only DECAY events
memory action=audit type="DECAY" sinceDays=7

Semantic Retrieval

Keyword matching runs first; when hits are insufficient, results are supplemented with a pure-JS TF-IDF + cosine implementation — no native modules, no external dependencies, fully offline. Semantic hits are marked as "semantic" in query output.

Configuration

// Plugin config (bundle or profile layer)
{
  halfLifeDays: 7,          // half-life in days for decay
  decayThreshold: 3,        // weight below this → archived
  consolidateThreshold: 3,  // references ≥ this → consolidate (+1 weight)
  weightCap: 20,            // consolidation weight cap (prevents runaway growth)
  hotTokenLimit: 5000,      // snapshot hot-section token budget
  maxQueryResults: 20,      // query result cap
  petEndpoint: null         // optional: local notification service URL (off by default)
}

Compatibility

• Node >= 22.19.0
• @deepseek-ai/dsh-* 0.1.0-rc.5 runtime (implemented against actual lib sources)

Troubleshooting (FAQ)

• Node version: requires Node >= 22.19.0; older versions may fail to load the plugin.
• DSH runtime compatibility: targets @deepseek-ai/dsh-* 0.1.0-rc.5 — check the version of the runtime you actually run.
• Memory directory issues: if writes fail, check read/write permissions on the memory root; if DSH_MEMORY_ROOT is set, it must point to an existing, writable directory.
• Native module conflicts: this plugin has no native dependencies — it is pure JS, so it cannot clash with native modules of other plugins.

Usage Scenarios

• Personal knowledge base, long-term maintenance: accumulate facts and decisions over time, query them later like a second brain; decay and archiving keep the store tidy without manual pruning.
• Project experience accumulation: lessons, habits and decisions live per-project in projects/<name>/, consolidating (weight grows) as topics are referenced repeatedly.
• Cross-session preference memory: preferences are injected at every session start, pin important ones for stability, and let conflict arbitration keep preferences authoritative over behavior.

Contributing

• Report issues: open an issue with the DSH runtime version, Node version, and reproduction steps.
• Pull requests: fork the repository, make the change, add/update tests, and run npm test before submitting.
• Tests: run npm test (node:test). New behavior should ship with test coverage.

License

MIT