memee 2.2.0

Cross-model shared memory for AI agents. One canon. Every model, every project, every teammate.

Memee

You shouldn't have to teach it twice.

Every chat is a new intern. You teach them Monday. By Friday they've quit. Memee writes it on the wall. The next intern reads it. So does your teammate's. So does the next model.

Memee shares experience across projects, across agents, across models, and across the people on your team.

pipx install memee         # recommended for a CLI tool
# already installed? pipx upgrade memee

For teams and companies. This OSS release is single-user and self-hosted. If you want the same memory shared across a whole team — with cross-developer, cross-agent, cross-project, cross-model canon building into company-wide institutional knowledge — there's a paid Team edition at memee.eu. Same engine, plus SSO, audit log, and shared scope. Flat $49 / month for up to fifteen seats, or $12k / year Enterprise with SOC 2 and air-gap.

---

What Memee actually does

Three jobs. Executed relentlessly.

Records.

Every pattern, every decision, every near-miss. One turn at a time, across every agent on every project.

7-task A/B: time −71 %, iterations −65 %, mistakes 0.

Routes.

Not a dump. A briefing. At task start, the router picks the 5–7 memories the agent actually needs — inside a hard 500-token budget. Your CLAUDE.md grows forever. Memee doesn't.

Measured ~40 tokens per task against a ~2,160-token median baseline.

Scores.

A lesson earns trust by surviving. A second model family agrees: confidence ×1.3. A second project re-uses it: ×1.5. Earn both and it climbs the ladder — hypothesis, tested, validated, canon.

One canon. Four model families. Seventeen engines.

---

Install and first use — sixty seconds

pipx install memee
memee setup                       # writes MCP + hooks to ~/.claude/settings.json
memee pack install python-web     # day-one canon, 30 entries

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

# Ask the canon what it knows about a snippet.
memee why "eval(user_input)"

# Health check (rerank state, hooks, cache).
memee doctor

That's it. Memory lives in ~/.memee/memee.db. No account. Core read/write is fully local. Vector embeddings are optional — on by default via sentence-transformers, which fetches a ~80 MB model on first use. Set TRANSFORMERS_OFFLINE=1 to skip.

After memee setup your agent doesn't have to remember Memee exists. SessionStart and UserPromptSubmit hooks land a routed briefing into Claude Code's context automatically; Stop fires a post-task review. Opt out with memee setup --no-hooks if you want the plain MCP-only mode.

---

The architecture, on one page

Small engines on SQLite + FTS5 + a 384-dim embedding space.

Layer	Job
Router	Task-aware briefing. Budget-capped.
Quality gate	Validates, deduplicates, rates every incoming memory before it earns a row.
Confidence scoring	Adaptive. Cross-project ×1.5. Cross-model ×1.3. Both stacked ×1.95.
Lifecycle	hypothesis → tested → validated → canon → deprecated. Old advice ages out. Good advice gets promoted.
Dream mode	Nightly. Connects related memories, surfaces contradictions, elevates canon.
Propagation	A validated pattern auto-pushes to projects with matching stack or tags. Fix once. Benefit everywhere.
Review	git diff | memee review - scans a changeset against known anti-patterns. Institutional memory enters code review.
CMAM bridge	Push canon to Anthropic's Managed Agents Memory at /mnt/memory/. Claude sees canon on turn one — no MCP round-trip.

Deeper notes: CLAUDE.md. CMAM spec: docs/cmam.md. Review engine: docs/review-fixes.md.

---

The token math

Numbers below are internal simulations and measured benchmarks, not independent third-party evaluations. Treat them as suggestive, not conclusive.

The thing Memee saves isn't the first page. It's the slope.

• Without Memee, median: ~2,160 tokens per turn. That's a CLAUDE.md / AGENTS.md across 27 popular OSS repos (langchain, vercel/ai, prisma, zed, openai/codex, and others), sampled via gh api. Claude Code and Cursor load it in full on every session.
• Without Memee, grown teams: 6k–15k. p95 of the sample hits 9,600. One published outlier reached 42,000.
• With Memee: 500-token cap, measured average ~40 tokens per briefing (min 18, max 67 across 10 task queries on a 500-pattern corpus).
• So the saving, honestly: ≥77 % at median. ≥95 % at 10k-grown teams. ≥99 % at the 42k outlier. And unlike CLAUDE.md, it's bounded. Your library grows. Per-turn context doesn't.

Reproduce locally:

memee benchmark          # OrgMemEval v1.0
pytest tests/ -v         # full suite

Full methodology + per-repo file sizes: docs/benchmarks.md.

---

Benchmarks

• OrgMemEval v1.0: 92.3 % across propagation, avoidance, maturity, onboarding, recovery, calibration, synthesis. Competitors on the same scenarios: MemPalace 0.9, Letta 1.3, Zep 2.3, Mem0 3.5 (the closest). v2.0.0 retired the autoresearch scenario alongside the engine.
• 7-task A/B (with / without Memee): time −71 %, iterations −65 %, quality 56 % → 93 %, ROI ≈ 10.7× at the $49 / month Team tier.
• GigaCorp simulation, 100 projects, 100 agents, 18 months: incidents 12/mo → 3/mo, annual ROI ≈ 3× at the same flat Team tier.
• Retrieval: 207-query × 255-memory eval harness with 7 difficulty clusters. BM25-only baseline nDCG@10 = 0.7273. With the optional cross-encoder rerank (MEMEE_RERANK_MODEL=cross-encoder/ms-marco- MiniLM-L-6-v2, pip install memee[rerank]): nDCG@10 = 0.7628 (+0.0355, p=0.0002). Run python -m tests.retrieval_eval to reproduce.

---

Using it with Claude, GPT, Gemini

An MCP server with 19 tools ships with the install. 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 tour:

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

Flat per team. Same engine in every tier.

	Free	Team	Enterprise
	$0 forever · MIT	$49 / month flat — up to 15 seats, annual	from $12k / year — unlimited seats
For	Solo developers. Self-hosted. Full engine, local scope.	Teams that want shared memory, SSO, and an audit trail.	Regulated industries, air-gap, SOC 2.
Stack	Router, quality gate, dream mode, CMAM sync, all 4 model families	Everything in Free + team/org scope with promotion workflows, SSO (SAML / OIDC), RBAC, audit log export, Postgres / Turso backend, multi-agent dashboard, 24h SLA	Everything in Team + SOC 2 Type II, DPA, SCIM, on-prem license key, dedicated CSM, 4h SLA, custom MCP integrations

Between fifteen and a hundred seats, and no SOC 2 needed? Email [e-mail] for a custom Growth plan.

Memee is memory, not model. Value scales sublinearly with headcount — one canon serves the whole team — so pricing is flat, not per-seat.

---

Memory, compared

Memory products cluster around two shapes: per-user context (ChatGPT, Hermes) or per-agent state (Letta, Mem0, Zep). Memee is the only one built per-organisation, with confidence scoring, cross-project propagation, and a flat-fee team tier.

Product	Cross-model	Cross-project	Cross-team	Auto-capture	Aging / lifecycle	Licence
Memee	Claude, GPT, Gemini, Llama	Yes. Propagation engine	Yes. Team / org scope	Both. Quality gate	Confidence + 60-day archive	MIT (OSS) + EULA (Team)
Claude Skills	Open standard. Ported to GPT, Gemini	Per-folder. Manual reuse	No. Static files	No. Author-written	None. Files don't decay	Free with Claude
Mem0	Claude, GPT, Gemini, local	Not first-class	Cloud tier	Auto-extract from chat	Not documented	Apache 2.0 + cloud from $19
Letta	Multi-model	Per-agent state	Not advertised	Auto via agent loop	Three-tier virtual memory	Apache 2.0 + cloud from $20
Zep	Multi-model	Not a documented primitive	Roles via API	Auto from episodes	Temporal graph	Graphiti Apache 2.0. Cloud from $25
ChatGPT Memory	Vendor-locked to ChatGPT	Project containers. Isolated	No. Per-user only	Both	Not published	Bundled with Plus / Team
Hermes Agent	Provider routing. Multi-LLM	Single-server posture	No	Both. Agent-curated	Not documented	MIT

Sources, verified April 2026: anthropic.com/news/skills, mem0.ai/pricing, letta.com/pricing, getzep.com, openai.com/memory, hermes-agent.nousresearch.com.

---

Three fair questions

From the launch thread. Answered without flinching.

Doesn't an MCP server just add to my context window?

It would, if it dumped. Memee doesn't. The router has a hard 500-token cap and lands at ~40 tokens on average per briefing. Measured, not aspirational.

Compare that to the file you already ship: median CLAUDE.md / AGENTS.md across 27 popular OSS repos is ~2,160 tokens, loaded in full on every session and ridden along on every turn. That file only grows. Memee is the cure for context bloat, not a new source of it.

Don't Claude Skills already do this?

Skills are good. They're also Claude-native, hand-authored, static, and per-folder. Memee is none of those things.

It writes from the conversation, not from your hand. It scores what it captures (cross-model ×1.3, cross-project ×1.5). It ages out what stops earning its keep on the lifecycle ladder, hypothesis → tested → validated → canon → deprecated. The same canon answers GPT in Cursor, Gemini in Continue, and Llama wherever you run it. Skills are a folder. Memee is a memory.

How is this different from Mem0, Letta, Zep, ChatGPT memory?

Most memory projects remember conversations: you talked about X last Tuesday, here it is again. That's chat history with a vector index.

Memee earns canon. A claim arrives at 0.5 confidence and goes nowhere until a second model family agrees and a second project re-uses it. On OrgMemEval v1.0 that capability gap shows up as 92.3 % against a competitor baseline of ~2 %. Not because the others are bad. Because they aren't built for the job.

Conversation memory remembers what was said. Institutional memory remembers what was learned.

---

Contributing

PRs welcome. Before opening a large one, a short issue describing the direction saves everyone a round-trip.

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

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

---

License

Memee core is MIT. The optional memee-team package is proprietary, distributed under a separate commercial EULA. See memee.eu for the terms.

---

Built by people who stopped teaching the same lesson to every new agent.