dsh-biomemory · Biomimetic Memory for DeepSeek Harness

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.

v0.5.2 (2026-08-20) — editable memories + conflict surfacing:

• Edit entries in place: memory action=update fp="..." text="..." (tool), /memory edit <fp> <new text> (command) and an Edit button on the Knowledge tab — metadata (pin/weight/layer) is preserved, the stale vector is cleared, and an UPDATE audit event is recorded; duplicates are rejected.
• Conflict surfacing: behavior memories that clash with user preferences are pinned to the top of memory action=list, the Knowledge tab (red badge + left border) and the frozen session snapshot (marked [冲突]) — so conflicts float up for you to judge and fix, instead of being silently auto-arbitrated away. Search results keep relevance order but still carry the conflict marker.
• Conflict arbitration in the Reflect tab: conflicts are listed at the top of the reflect view, each with inline Edit (large auto-growing editor) and Delete buttons, plus an Undo bar after deletion.
• Single-entry rollback: memory action=restore fp="..." (tool), /memory undo <fp> (command), POST /entries/restore (API) restore a deleted entry from the newest backup DB — metadata preserved, RESTORE audit event recorded.
• Reflect data source fixed to SQLite (was scanning the Markdown read-only backups — deleted entries resurrected on the next reflect).

v0.5.1 (2026-08-20) — pet bridge extracted:

• The DSH↔desktop-pet bridge (session/event state forwarding + approval panel) moved out to its own plugin dsh-whale-pet-bridge — biomemory now only handles memory (save notifications to the pet are kept).
• No functional changes to memory features.

v0.5 (2026-08-19) — SQLite + semantic retrieval:

• SQLite data layer (~/.dsh/biomemory/biomemory.db, Node 24 built-in node:sqlite, WAL mode, zero external deps) — L2/L3 structured entries + vector blobs + audit log
• Offline embedding model (bge-small-zh-v1.5, 512-dim, quantized ONNX ~24MB at ~/.dsh/models/) via transformers.js — pure JS, no native modules
• Three retrieval modes: exact (keyword) / semantic (vector) / hybrid (default, Reciprocal Rank Fusion per v0.5 design doc §3.4)
• Automatic Markdown migration: existing ~/.dsh/memory entries imported once on first boot (Markdown kept as read-only backup)
• Audit aggregation (P1-003): group by action / day / entry
• Dream checkpointing (P0-002): resume interrupted metabolism from the last checkpoint
• Graceful degradation: model unavailable → keyword retrieval only; memory features unaffected

Core features (unchanged from v0.4):

• memory tool: add / query / update / 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
• /memory command: list / query / add / edit / 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, cold archiving (status flag, never deleted)
• Memory pins: lock a memory so it never decays and always enters the snapshot

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 surfacing (v0.5.2): when behavior memory conflicts with preferences it is no longer silently halved — it is exempted from decay/archival, stays active, and floats to the top of listings and the session snapshot so you can judge and edit it. A CONFLICT audit event is recorded; once you edit it into agreement, normal metabolism resumes.
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.

Auto recall / auto save (v0.4.0)

Three automatic layers on top of explicit calls:

1. Approval fallback: important memories normally require approval; when approval is unavailable (policy never / service missing), they are saved automatically per approvalFallback (default auto), audited as [降级]. Switch to deny in settings to stay fail-closed.
2. Auto consolidation (use-it-or-lose-it): every keyword query/recall hit bumps hits+1 and writes back — memories that get recalled often decay slower (audit RECALL).
3. Auto dream/reflect: autoDreamDays (default 7) and autoReflectDays (default 3) run metabolism/reflection at startup when older than the interval; 0 disables. Audit AUTO-DREAM / AUTO-REFLECT.

Deep reflection (Reflect, v0.4.0)

/memory reflect (or memory action=reflect, settings tab) — a purely local, LLM-free periodic summary:

1. Topic clustering: all entries clustered by TF cosine similarity (≥0.25) to surface recurring topics;
2. Trend stats: writes in the last 7 days vs the previous week (rising / steady);
3. Conflict alerts: behavior memories that clash with preferences;
4. Forget candidates: low-weight entries worth reviewing.

Reports are written to longterm/reflections/<timestamp>.md; --dry-run previews without writing.

Knowledge page (v0.4.0)

The settings page gains a Knowledge tab: full-text/semantic search, layer filter, per-entry weight/hits/time/pin display, one-click pin/unpin, edit (inline textarea) and safe removal (backed up first, restorable). Conflicting behavior entries float to the top with a red badge. Web API: GET /biomemory/api/entries, POST /biomemory/api/entries/pin|unpin|update|remove, POST /biomemory/api/reflect.

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