pi-harnmem0.4.1

pi-harnmem

Long-term, plain-file memory for pi — the pi adapter for harnmem.

Memory is stored as Markdown in ~/.harnmem (no service, no database, no local embedding model). The extension injects a small, byte-stable memory index into the session system prompt and exposes 5 tools so the model can write, search, read, forget and inspect memory across sessions and projects.

Requires pi >= 1.3.0 (the peer range accepts 1.4.x). A DSH adapter with the same feature set also exists: dsh-harnmem.

Install

pi install npm:pi-harnmem
# -l / --local installs into the current project's settings instead of user settings

Install into pi (local path)

For development against a repo checkout:

pnpm install && pnpm -r build
pi install ./packages/pi-harnmem

Verified syntax against pi install --help (pi 1.3.0). The full install flow is not covered by the repo's automated checks — if it fails, use the dev loading path below.

Dev loading (no install)

pi -e ./packages/pi-harnmem/dist/index.js

Verified end-to-end against a real pi 1.3.0 session: empty-store load, write, cross-session recall, byte-stable injection, read-only children.

Quick start

In a pi session:

You: remember this project uses pnpm — never npm install.
pi:  → memory_write(...) → stored mem_01J8... (NEW)

Next session, the hot index is injected automatically:

## HarnMem memory (cross-harness; stable snapshot)

- 📌 [user] Dependencies in this project use pnpm — never npm install (avoids lockfile drift) (mem_01J8...)

<!-- SSOT: ~/.harnmem. Use memory_search for details; never invent memory ids. ... -->

An empty store injects nothing (0 bytes).

Configuration

Settings are read (lowest → highest priority) from: built-in defaults → <home>/config.json → <cwd>/.harnmem.json → <agentDir>/harnmem.json → environment variables. agentDir is PI_CODING_AGENT_DIR or ~/.pi/agent. JSONC comments are allowed; a malformed file produces a warning, never a crash.

<agentDir>/harnmem.json:

{
  "enabled": true,          // master switch
  "inject": true,           // inject hot index at session start
  "readOnly": false,        // block all writes
  "statusLine": true,       // footer line "harnmem: N entries · ~T tokens · <sha>"; false hides it
  "autoExtract": false,     // background extraction on session quit (opt-in)
  "home": "~/harnmem-home",  // default: ~/.harnmem
  "maxInjectTokens": 1200,  // injection budget (800..2000 recommended)
  "maxSearchTokens": 1200,
  "model": {                // LLM used for extraction/consolidation
    "provider": "<provider>",
    "id": "<model>",
    "reasoning": "low",
    "timeoutMs": 60000
  },
  "log": "errors",          // off | errors | on  → <home>/logs/pi.jsonl
  "toolPrefix": "memory_",  // model-facing tool prefix; use "harnmem_" when pi-memory is also installed
  "debug": false
}

Environment overrides: HARNMEM_HOME, HARNMEM_DISABLED=1, HARNMEM_READONLY=1, HARNMEM_INJECT=1|0, HARNMEM_STATUS_LINE=1|0, HARNMEM_AUTO_EXTRACT=1, HARNMEM_DEBUG=1.

Clean footer. The one-line harnmem: 22 entries · ~1061 tokens · <sha> status is informational; set "statusLine": false in <agentDir>/harnmem.json (or HARNMEM_STATUS_LINE=0) to hide it. Memory injection, tools and extraction are unaffected; /memory status and the memory_status tool still report the snapshot.

Coexisting with pi-memory. Both extensions otherwise register the same tool names (memory_write, memory_search, …), and pi silently keeps one of the duplicates. Set "toolPrefix": "harnmem_" in <agentDir>/harnmem.json to expose harnmem_write, harnmem_search, harnmem_read, harnmem_forget, harnmem_status instead. The /memory command keeps its name (pi-memory ships no conflicting command).

Tools (model-facing)

Tool	Parameters	Notes
memory_write	type (user/feedback/project/reference), title, description, content, category, scope (auto/user/project), importance (0–10), confidence (0–1), pinned, tags, entities, links, supersedes, evidence	Returns one line: stored … (NEW) / already exists … (SKIP) / merged into … (MERGE) / superseded … → … / linked to existing … (LINK)
memory_search	query, limit (8), scope, project, include_archived, include_history, as_of, entity, debug	BM25F + synonyms + link diffusion; abstains when not confident
memory_read	id, as_of, include_history	Frontmatter summary + body + backlinks
memory_forget	id, hard	Default moves to trash/ (30-day grace); hard is permanent
memory_status	—	Shards, index state, health, extraction counters

memory_write marks the snapshot dirty; every successful write refreshes the injected index in the next step.

Tool names follow the toolPrefix setting (memory_ by default, harnmem_ in the recommended coexistence setup); the table above uses the default names.

Commands (human-facing)

/memory status
/memory search <q> [--limit N]
/memory read <id> [--history]
/memory forget <id> [--hard]
/memory restore <id>
/memory consolidate [--dry-run] [--mechanical]
/memory help

How it works

• Stable snapshot (Option-P): the context block is rebuilt only on session start, before compaction, after a successful write, and on day rollover. Within a session it is byte-identical, so the provider prompt cache is not invalidated. Rebuild cost is well under 20 ms for typical stores.
• Injection budget: maxInjectTokens (default 1200). Pinned entries first, then user/project/digest sections; whole entries only. Body text is never injected — the model calls memory_read for detail.
• Read-only children: subagent sessions (child markers such as PI_SUBAGENT_CHILD) get no LLM and cannot write: memory_write failed: HarnMem is read-only.
• Session end: the transcript (last ≤200 messages, assistant truncated) is journaled; when autoExtract is on, extraction runs at quit with an 8 s hard timeout, ≥3 user messages, and is skipped if the model wrote a memory itself that session. Extracted candidates below confidence 0.5 are dropped; at most 6 are stored.
• Custom extraction prompt: <home>/prompts/extract.md is loaded when present and parsed into ## system / ## user sections; placeholders such as {{observation_date}}, {{existing_memories}} and {{transcript}} are substituted into the user template. Missing files or sections fall back to built-in English defaults, and editing the file takes effect on the next extraction without restarting.

Storage

~/.harnmem/
├── VERSION
├── config.json            # optional engine config (budget/retrieval/write/governance)
├── projects.json          # project registry (git remote / path → projectId)
├── memories/user/         # MEMORY.md index + mem_*.md records
├── memories/projects/<projectId>/
├── index/                 # manifest.json + <shard>.json (rebuildable)
├── digests/<shard>/       # optional topic digests
├── sessions/<date>/       # transcript journals for extraction
├── state/{seq,audit.jsonl,cursors.json}
├── trash/                 # 30-day recycle bin
└── logs/pi.jsonl

Plain files are the source of truth: back up or sync ~/.harnmem directly. The index can always be rebuilt from the Markdown records.

Troubleshooting

Symptom	Fix
No memory block in context	Empty store (normal); check /memory status; check enabled/inject/HARNMEM_DISABLED
Write says HarnMem is read-only	Subagent session, readOnly: true, or HARNMEM_READONLY=1
index: ... stale=true right after a write	Expected: index rebuilds at next session start
Extraction reports failed	Check LLM availability and <home>/logs/pi.jsonl
Slow exit (~8 s)	Bounded extraction wait; disable autoExtract if not wanted
Provider 400 with --thinking max	Use --thinking low (known provider quirk)

Development

pnpm --filter @harnmem/core build
pnpm --filter pi-harnmem build     # tsc; 44 files emitted to dist/
pnpm test                          # core test suite (52 tests)

Source layout: src/index.ts (event wiring), state.ts (engine lifecycle), context.ts (snapshot manager), tools.ts, command.ts, llm.ts (pi model bridge), journal.ts, log.ts, subagent.ts, config.ts, format.ts.

Limitations

• pi only for this package; DSH uses dsh-harnmem, and Claude Code has a design doc only (docs/design/09).
• No automated tests for this package; behavior is covered by manual end-to-end runs against pi 1.3.0 (real subagent sessions not yet exercised).
• autoExtract defaults to off and only runs on normal quit.

License

MIT