memee 1.0.4

Institutional memory for AI agent teams: multi-model, cross-project, token-budgeted.

Memee

Institutional memory for AI agent teams. Your agents stop re-solving problems.

Memee sits between your agents and the work. It records what worked, flags what didn't, and hands each agent only the 5–7 memories it actually needs for the task in front of it — instead of re-stuffing 500 patterns into every prompt.

pipx install memee            # recommended for a CLI tool
# or:
python3 -m pip install memee  # if you don't have pipx

---

Why Memee

• Your agents stop repeating mistakes. One agent hits a bug, records it, every other agent (in every other project) is warned before it happens again.
• Send 500 tokens, not 14,550. A smart router picks the memories relevant to the current task, inside a configurable token budget. Internal sim: 96% reduction per task.
• Every model reads from one canon. Claude, GPT, Gemini, Llama all record into, and pull from, the same memory. Cross-model agreement raises confidence; single-model claims stay provisional.

Install and first use (60 seconds)

pipx install memee
memee setup

# Record a pattern you just learned
memee record pattern "retry with jitter" \
  --tags reliability,http \
  -c "Always use exponential backoff with jitter; capped at 30s; retry only idempotent verbs."

# Search it back
memee search "retry"

# Health check + auto-wire Claude Code / MCP config
memee doctor

That's it. Memory lives in ~/.memee/memee.db. No account. Core read/write is fully local; optional vector embeddings use sentence-transformers, which fetches a ~80 MB model from HuggingFace on first use (skipped when TRANSFORMERS_OFFLINE=1 is set).

What it actually does

Memee is a stack of small engines sitting on SQLite + FTS5 + sentence-transformer embeddings.

Layer	What it does	Why you care
Router	Task-aware briefing. Picks 5–7 memories within a token budget.	Agents get signal, not a dump.
Quality gate	Validates, dedupes, rates every incoming memory.	Junk doesn't survive the first write.
Confidence scoring	Adaptive: cross-project ×1.5, cross-model ×1.3, combined ×1.95.	Patterns earn trust across evidence, not by author claim.
Lifecycle	hypothesis → tested → validated → canon → deprecated.	Old advice ages out; good advice gets promoted.
Dream mode	Nightly: connect related memories, surface contradictions, promote canon.	Knowledge compounds while you sleep.
Propagation	A validated pattern auto-pushes to other projects with matching stack/tags.	Fix once, benefit everywhere.
Review	`git diff	memee review -` scans changes against anti-patterns.
CMAM bridge	Push canon to Anthropic's Managed Agents Memory at /mnt/memory/.	Claude sessions see canon on turn 1, no MCP roundtrip.

Deeper architecture doc: CLAUDE.md. CMAM specifics: docs/cmam.md. Review engine: docs/review-fixes.md.

Benchmarks

All numbers below are internal simulations and benchmarks, not independent third-party evaluations. They describe system behaviour under synthetic workloads. Treat them as suggestive, not conclusive.

• Token savings per task: 14,550 → 500 (–96%) in the router micro-benchmark.
• OrgMemEval v1.0: 92.4/100, across propagation, avoidance, maturity, onboarding, recovery, calibration, synthesis, and research. Competitor baselines on the same scenarios range from 0.9/100 (MemPalace) to 3.5/100 (Mem0, the closest); see docs/benchmarks.md.
• 7-task A/B (with vs. without Memee): time –71%, iterations –65%, quality 56% → 93%, impact-DB ROI ≈ 10.7× at the current $49/mo Team tier. GigaCorp sim (100 projects, 100 agents, 18 months): incidents 12/mo → 3/mo, annual ROI ≈ 3× at the same flat Team tier.
• Retrieval: hit@1 = 100% on a 12-memory routing benchmark after the recent ranking fix.

Reproduce locally:

memee benchmark          # OrgMemEval v1.0
pytest tests/ -v         # full suite, ~60s

Using it with Claude, GPT, Gemini

Memee ships an MCP server with 24 tools. Drop this into ~/.claude/settings.json (or the Cursor / Continue / any MCP-capable client equivalent):

{
  "mcpServers": {
    "memee": { "command": "memee", "args": ["serve"] }
  }
}

Memee auto-detects the caller's model family from MEMEE_MODEL, ANTHROPIC_MODEL, or OPENAI_MODEL and tags every write with source_model. That's how confidence scoring knows when Claude and Gemini agree — and when they don't.

Quick CLI:

memee brief --task "write unit tests"   # PUSH: routed briefing
memee check "about to add eval() here"  # PULL: anti-pattern check
memee propagate                          # cross-project diffusion
memee dream                              # nightly: connect, contradict, promote
memee cmam sync                          # push canon to /mnt/memory/ for Claude

Pricing

Memee (this repo) is MIT-licensed and free. It's a single-user product: your memory, your machine, no account.

If you need multi-user scope (personal / team / org), SSO, audit log, seat management, or a hosted control plane, install memee-team — a paid proprietary package from memee.eu. Pricing is flat, not per-seat: $49/month for a team of up to 15, from $12k/year for Enterprise with SOC 2, SCIM, and air-gap. It plugs into the same engine; no re-install, no data migration.

Contributing

PRs are welcome. Before opening a big one, a short issue describing the direction saves everyone time.

pip install -e ".[dev]"
pytest tests/ -v

Style: type hints, docstrings in English, 100-char lines, ruff clean. New engines live in src/memee/; every new behaviour wants a test in tests/.

License

Memee core is released under the MIT License. The optional memee-team package is proprietary and distributed under a separate commercial EULA; see memee.eu for terms.