compound-memory

English | 简体中文

Local-first shared memory for multiple AI agents — plain Markdown files that compound in value as they are used. Memory lives on your disk as frontmatter-annotated Markdown, gets stronger with every confirmed use, decays into a revivable archive when neglected, and auto-commits to a local git history on every write.

Why

Every agent session starts from zero: preferences get re-asked, project conventions get re-discovered, the same pitfall gets hit twice. compound-memory gives all your agents one shared store:

• Local-first — nothing leaves your machine; memories are human-readable Markdown files, not rows in an opaque database.
• MCP-native — exactly 5 tools (memory_write / memory_search / memory_get / memory_link / memory_feedback) as the single read-write boundary; works with any MCP host (Claude Code, ZCode, WorkBuddy, …), plus a full CLI for operations.
• Compounding — confirmed usage raises confidence, related memories are recalled as neighbors, validation from a different host counts as independent evidence, and distillation merges many raw memories into fewer, denser ones.
• Multi-agent by design — a _shared namespace everyone reads, plus agent-* private namespaces each host owns; cross-host validation is tracked per host.
• Optional semantic recall — vector search via sqlite-vec + BGE embeddings, with automatic graceful fallback to pure lexical search when unavailable.

Quick Start

For AI agents

Paste this one-liner into your coding agent (Claude Code, Cursor, ZCode, …) and let it do the rest:

Set up compound-memory (https://github.com/chinwe/compound-memory) — a local-first multi-agent shared memory (MCP server + CLI) — on this machine: install it (`uv tool install compound-memory`, or clone the repo and `uv sync --extra dev`), initialize the store (`compound-memory init`, defaults to ~/.agents/memory), register its stdio MCP server in this host's MCP config — command `compound-memory-server` (PyPI install) or `uvx --from compound-memory compound-memory-server`, env `COMPOUND_MEMORY_ROOT=$HOME/.agents/memory` and `COMPOUND_MEMORY_AGENT_ID=agent-<your-host-id>` — then verify by calling `memory_search` and expecting a `{"hits": [...]}` response; if the host needs a restart to load MCP servers, tell me. Host-specific configs and the usage protocol: docs/agent-integration.md in the repo.

For humans

1. Install

Python ≥ 3.11. Either route works:

# Route A: clone the repo (uv-managed; same path the MCP config uses)
git clone https://github.com/chinwe/compound-memory.git
cd compound-memory && uv sync --extra dev

# Route B: install from PyPI (no clone needed)
uv tool install compound-memory   # or: pip install compound-memory

2. Initialize your store

Defaults to ~/.agents/memory; override with the COMPOUND_MEMORY_ROOT env var. A leading ~ or $HOME is expanded; a blank override falls back to the default.

uv run compound-memory init

3. Wire it into your MCP host (recommended)

This lets your everyday agents read/write the shared store automatically:

{
  "mcpServers": {
    "compound-memory": {
      "type": "stdio",
      "command": "uv",
      "args": ["run", "--directory", "<repo>", "compound-memory-server"],
      "env": {
        "COMPOUND_MEMORY_ROOT": "$HOME/.agents/memory",
        "COMPOUND_MEMORY_AGENT_ID": "agent-<your-host-id>"
      }
    }
  }
}

Installed from PyPI? Swap command/args for uvx + ["--from", "compound-memory", "compound-memory-server"] — no repo clone needed. COMPOUND_MEMORY_AGENT_ID is required for any multi-agent store: the process identity wins over whatever the model claims, so a forged source cannot write another host's namespace or inflate cross-host validation. Leave it unset only for a single trusted caller on a private machine — stats then reports attestation.warning, and the MCP server logs the same warning at startup. Without it, identity is self-reported and can be forged.

Verify: ask your agent to call memory_search (any keyword) — a {"hits": [...]} response means you're connected. Or run uv run compound-memory stats from the CLI.

4. Next step

Inject the usage protocol from skills/compound-memory/SKILL.md into your host (the search → feedback → distill loop), per docs/agent-integration.md §6.

Demo

One full loop: write → search → feedback (with cross-host first-validation bonus) → store health. Real output from v0.4.0, long payloads trimmed:

uv run compound-memory init

{ "ok": true, "root": "~/.agents/memory" }

Two different hosts each write one stable fact (new memories start at confidence 0.5, uses 0):

uv run compound-memory write \
  "Deploy serverless functions on this platform times out at 10s — keep handlers under that budget" \
  fact agent-claude --key vercel-timeout

{
  "id": "20261007_86adf1",
  "ns": "_shared",
  "type": "fact",
  "source": "agent-claude",
  "content": "Deploy serverless functions on this platform times out at 10s — keep handlers under that budget",
  "confidence": 0.5,
  "uses": 0,
  "key": "vercel-timeout",
  "validated_by": []
  ...
}

uv run compound-memory write \
  "User prefers concise replies with tables and code examples" \
  fact agent-zcode --key user-style

Search ranks by score (--explain attaches per-hit ranking components for debugging):

uv run compound-memory search "serverless timeout"

[
  {
    "id": "20261007_86adf1", "score": 1.0292, "similarity": 1.0,
    "type": "fact", "source": "agent-claude",
    "content": "Deploy serverless functions on this platform times out at 10s — keep handlers under that budget",
    "neighbors": []
  },
  {
    "id": "20261007_6a0c0c", "score": 0.5211, "similarity": 0.4919,
    "type": "fact", "source": "agent-zcode",
    "content": "User prefers concise replies with tables and code examples",
    "neighbors": []
  }
]

A different host used this memory and reported it back with the default success outcome — uses +1, conf +0.1; and since the reporter agent-workbuddy ≠ source agent-claude, the first cross-host validation adds another +0.15:

uv run compound-memory feedback 20261007_86adf1 agent-workbuddy

{
  "id": "20261007_86adf1",
  "confidence": 0.75,
  "uses": 1,
  "last_used": "2026-10-07",
  "validated_by": ["agent-workbuddy"],
  "evidence": {
    "success_count": 1, "failure_count": 0, "contradiction_count": 0,
    "last_verified": "2026-10-07",
    "recent": [{ "date": "2026-10-07", "agent": "agent-workbuddy", "outcome": "success" }]
  }
  ...
}

Store health at a glance (fixed-bucket histograms, liveness, distillation yield):

uv run compound-memory stats

{
  "total": 2, "archived": 0, "active": 2,
  "avg_confidence": 0.625,
  "by_type": { "fact": 2 },
  "by_ns": { "_shared": 2 },
  "review_queue_entries": 0,
  "uses_histogram": { "0": 1, "1-2": 1, "3-5": 0, "6-9": 0, "10+": 0 },
  "confidence_histogram": { "<0.3": 0, "0.3-0.6": 1, "0.6-0.8": 1, "0.8-1.0": 0 },
  "recent_feedback_7d": 1, "cross_validated": 0,
  "distilled_total": 0, "distilled_recent_7d": 0
}

Current stats also includes retrieval (mode: vector or lexical, plus reason when the vector channel is off) and attestation (attested, plus a warning when COMPOUND_MEMORY_AGENT_ID is unset). memory_search carries the same retrieval object. CLI search stays a bare array.

Three things to notice:

• New memories start at confidence 0.5 and move on evidence — feedback carries an outcome: success raises it, failure lowers it (floor 0.05), contradiction freezes it into the review queue, obsolete archives immediately.
• First validation from a different host earns an independent bonus (once per host per memory), with validated_by / evidence trails — confidence is evidence of correctness, not popularity.
• Hits embed one-hop neighbors automatically (empty here — no links yet; memory_link creates bidirectional links that get recalled for free).

How compounding works

Scoring (weights are the W_* constants in src/compound_memory/scoring.py): 0.70·similarity + 0.15·confidence + 0.10·recency(0.5+0.5·e^(−Δt/τ)) + 0.05·type weight. With the vector channel enabled, ranking switches to RRF fusion with an ε=0.04 prior tie-break (see the spec, "index as cache").

The 5 MCP tools

Tool descriptions embed the protocol rules themselves, so agents keep the loop intact even without host-side rules injected. Full parameter reference: docs/agent-integration.md.

CLI

uv sync --extra dev              # first clone: build .venv (later `uv run` reuses it)

uv run compound-memory init               # initialize an empty store
uv run compound-memory write "Vercel Serverless has a 10s timeout" episode agent-workbuddy
uv run compound-memory search "Vercel timeout"   # hits embed one-hop neighbors (limit 3, --no-neighbors to disable)
uv run compound-memory feedback <id> agent-claude
uv run compound-memory decay          # run from cron
uv run compound-memory revive <id>    # revive an archived memory
uv run compound-memory distill-plan   # distillation candidates: merge_with (same-key strong) + possible_dup_of (BM25 weak) + promotion_candidate (high-activity episodes)
uv run compound-memory distill-apply "the merged insight" insight agent-workbuddy --sources <id1>,<id2>  # atomic: product (links, origin=distillation) + source archival, one commit
uv run compound-memory stats            # health: uses/confidence buckets + liveness + distillation yield
uv run compound-memory rebuild-index  # rebuild the search cache anytime
uv run compound-memory review-queue   # conflict queue (CLI-only entry)
uv run compound-memory git-log        # audit trail

More operations: explain <id> (confidence composition + evidence detail for one memory), forget <id> --agent <id> (terminal removal, ADR-0009), review-resolve (adjudicate conflicts), extract <transcript|dir> (deterministic session-transcript mining).

Architecture

Agent (MCP client / CLI)
  └─ memory_write | memory_search | memory_get | memory_link | memory_feedback
       └─ MemoryStore (~/.agents/memory)
            ├─ namespaces/_shared/{episode,fact,insight,skill,decision}/*.md   shared area
            ├─ namespaces/agent-*/...                                          private areas
            ├─ archive/...                                                     decayed archive (revivable)
            ├─ index/tokens.json                                               rebuildable lexical cache
            ├─ index/vectors.db                                                rebuildable vector cache (optional)
            ├─ review-queue.md                                                 fact/insight/decision conflict queue
            └─ .git/                                                           auto-commit on every write

Scheduled distillation prep (launchd / cron / systemd)

Per ADR 0001, the deterministic prep runs on a schedule while judgment (summarizing / merging) stays with the calling agent. Every day at 09:00 the candidate list lands in <root>/distill/last-plan.json. Pick one scheduler — launchd (macOS standard, catches up after sleep), systemd user timer (Persistent=true, same catch-up), or cron (most portable, no catch-up) — all three drive the same platform-neutral scripts/distill-prepare.sh. Ready-made templates with copy-paste instructions: scripts/com.compound-memory.distill-prepare.plist.tmpl (launchd), scripts/compound-memory-distill-prepare.{service,timer}.example (systemd), and the Chinese README for cron. The script runs set -eu: any failure exits non-zero (visible via launchctl list / systemctl --user list-timers / cron mail, log at distill/prepare.log). distill/ is a runtime artifact directory (auto-gitignored) — no commit noise; only distill-apply after agent judgment lands one atomic commit.

Documentation

• CHANGELOG.md — release notes (0.5.x Windows support and later)
• docs/specs/0001-compound-memory-spec.md — design spec
• docs/agent-integration.md — per-host MCP configs + the unified usage protocol (Chinese)
• docs/adr/ — architecture decision records
• CONTEXT.md — glossary (Chinese)
• skills/compound-memory/SKILL.md — usage rules for hosts (Chinese)

Development

uv run pytest tests/ -q     # full suite (MCP tool boundary + distillation + lifecycle/index/CLI + input defense)
uv run mypy src/compound_memory/

Test seams: the MCP tool boundary via in-process mcp.Client(server) (no subprocess) plus unit tests for core modules (scoring / index / store ops). CI runs tests, type checks, and a pure-wheel install smoke on Linux and Windows across Python 3.11/3.12/3.13.

Platform support

Python ≥ 3.11 on macOS, Linux, and Windows.

Release

PyPI versions are immutable and the tag must match pyproject.toml's version (the release workflow verifies this and fails loudly). Releases go through GitHub Actions + PyPI Trusted Publisher (OIDC, no token): push a tag like v0.1.0 and release.yml builds and publishes automatically.

MCP Registry name: mcp-name: io.github.chinwe/compound-memory