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.
CLAUDE.md filesThree jobs. No dashboards, no copilots, no magic.
When any agent figures something out (a retry rule, a PII edge case, a deploy gotcha), Memee writes it down, dates it, scores it. The next agent on the next project picks it up without being told.
Your CLAUDE.md reloads in full on every turn. Memee picks the 5–7 memories this task actually needs, inside a 500-token cap, and stops there. Median router output: 39 tokens. The cap holds as your library grows past a thousand entries.
Claude, GPT, Gemini, Llama all write to and read from the same memory. One model’s claim is provisional. A second family confirms: confidence ×1.3. A second project re-uses: ×1.5. Combined: ×1.95. Nothing reaches canon on a single voice.
Mark a policy authoritative. It surfaces under Pinned policies in scope: in every briefing whose tags overlap — ahead of search, regardless of similarity score. The org rule outranks the conversational hit. memee record --pin.
Numbers from a 27-repo sample of real public
CLAUDE.md / AGENTS.md files, not a
synthetic worst case. The hidden problem isn’t the
first page — it’s the slope.
CLAUDE.md / AGENTS.md across 27 popular OSS repos (langchain, vercel/ai, prisma, zed, openai/codex, …). Claude Code and Cursor load it in full on every session start. Grown teams hit 6k–15k; a published pathological case reached 42k.CLAUDE.md, it’s bounded — it doesn’t grow as your knowledge base does.Your CLAUDE.md grows forever. Memee doesn’t.
gh api. Median file size
translates to ~2,160 tokens, 2,500 mean, 9,600 at p95, one
published outlier at 42,000. Anthropic’s own docs
confirm CLAUDE.md is loaded on every session
and rides along with every turn.pytest tests/test_router.py::test_token_budget_respected.If you pay per token, cut it from your bill. If you pay a flat subscription (Claude Max, ChatGPT Plus / Pro, Claude for Business seats), the saved tokens buy you four other things.
Max x20 isn’t unlimited. You hit the 5-hour throttle. A 500-token briefing leaves your budget for the actual answer, not a context dump.
On a 207-query eval harness with adversarial lexical-gap queries, MRR = 0.87 with the optional cross-encoder reranker. In a 7-task A/B, iterations per task dropped −65 %. Real time saved, not imaginary dollars.
Claude Max session resets. Your agent forgot what you taught it an hour ago. Memee didn’t. One lesson, everywhere, forever. This part is independent of pricing tier.
Switch Claude Max to ChatGPT Team to Gemini. The canon stays. Flat subscriptions lock you to one vendor. Memee gives you a vendor-exit option.
Memee’s first GUI is one dot in your macOS menubar. It tells you Memee is awake. It tells you what it did. It writes nothing.
Status, last brief, last learn, totals. No notifications, no badges, no popups. The bar reads what the hooks wrote. memee bar install wires it as a LaunchAgent.
One rolling line: 3 mistakes avoided · 42 patterns applied · +2 canon this week. Zero counters drop out. Zero across the board hides the row.
A line appears when a newer PyPI release exists. Click opens the release notes. Hidden when you’re current. 24h cache, zero new network calls on the hot path.
A rule promoted two years ago is a candidate for being wrong today. Memee schedules its own re-checks. Per memory. Per half-life.
Every memory carries a predicted recall R(t) = 2^(-Δt/h). Validate it under load, the half-life stretches. Skip it, it shrinks. Anki’s scheduler, applied to canon.
When a wide-uncertainty canon row gets stale, Memee adds Re-checking canon: <title> to the briefing. A quiet flag, not a directive. Per-week, not per-prompt.
Canon used to require conf ≥ 0.85 and ten validations. v2.4.2 replaces both with the Sequential Probability Ratio Test at α = β = 0.05. One statistical dial; tighter Type-I.
Per-user context. Per-agent state. Or per-organisation memory. Pick the shape that fits your team.
| Product | Cross-model | Cross-project | Cross-team | Auto-capture | Aging / lifecycle | Re-checks itself | Licence |
|---|---|---|---|---|---|---|---|
| Memee | Claude, GPT, Gemini, Llama | Yes. Propagation engine | Yes. Team / org scope | Both. Quality gate | Confidence + 60-day archive | Yes. FSRS-light decay, SPRT promotion | 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 | No | Free with Claude |
| Mem0 | Claude, GPT, Gemini, local | Not first-class | Cloud tier | Auto-extract from chat | Not documented | No | Apache 2.0 + cloud from $19 |
| Letta | Multi-model | Per-agent state | Not advertised | Auto via agent loop | Three-tier virtual memory | No | Apache 2.0 + cloud from $20 |
| Zep | Multi-model | Not a documented primitive | Roles via API | Auto from episodes | Temporal graph | Temporal edges only | Graphiti Apache 2.0. Cloud from $25 |
| ChatGPT Memory | Vendor-locked to ChatGPT | Project containers. Isolated | No. Per-user only | Both | Not published | Not published | Bundled with Plus / Team |
| Hermes Agent | Provider routing. Multi-LLM | Single-server posture | No | Both. Agent-curated | Not documented | No | MIT |
Sources, verified April 2026: anthropic.com/news/skills, mem0.ai/pricing, letta.com/pricing, getzep.com, openai.com/memory, hermes-agent.nousresearch.com.
Five hand-curated .memee packs ship today. 306 sourced rules, signed, idempotent. One command to import. Day-one users get canon before the first agent session.
Async correctness, observability, deployment, security canon every Python service needs from day one. memee pack install python-web.
TypeScript strict, Tailwind, TanStack Query, React Router v6 loaders. Animation perf, prefers-reduced-motion, iOS 100vh fix. memee pack install react-vite.
Lethal-trifecta read+write split, supply-chain pinning, prompt injection in tool descriptions, stdio framing, the 15-tool ceiling. memee pack install mcp-server-canon.
RFC 9110 + 7807/9457 + 8725 + 9700 + OAuth 2.1 distilled. JWT BCP, OAuth pitfalls, problem+json errors, ETag concurrency. memee pack install http-api-canon.
Fabricated paths, premature done, scope explosion on small fixes, sycophancy under pushback, the lethal trifecta. Every entry cites primary literature. memee pack install agent-discipline.
Two more in the pipeline. SwiftUI state ownership, gesture/offset structural bugs. Postgres index strategy, EXPLAIN reading, lock contention. Open a PR.
From the replies. Answered without flinching.
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: a 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.
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.
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 96.3 % 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.
After memee setup, yes. Hooks land in your Claude Code settings: SessionStart fires a routed briefing into the agent’s context. UserPromptSubmit fires a task-aware brief on every prompt. Stop fires a post-task review. Three lines in your settings.json. You don’t write them; memee setup writes them. You don’t call memory_search; the hook called it.
Opt out with memee setup --no-hooks and use Memee as a plain MCP tool. Most don’t.
The agent doesn’t have to remember Memee exists. The runtime injects what it needs.
A hundred agents across a hundred projects. Eighteen months. Incidents fell from twelve a month to three. That’s the whole pitch.Internal simulation,
test_gigacorp, marked as such.
Flat per team. Same engine in every tier.
Free ships the full engine. Team layers on multi-user scope, SSO, and an audit log. Enterprise adds SOC 2, on-prem, and a 4h SLA. Between fifteen and a hundred seats and no SOC 2 needed? info@memee.eu handles Growth plans by hand.
Three lanes. Pick the one for you.
pipx install memee && memee setup memee pack install python-web # already installed? pipx upgrade memee
The engine, on your machine, by tea-time. Bring your own keys. Pick a pack — five ship today — and your first session starts already knowing.
No pipx?brew install pipx, or fall back to python3 -m pip install memee.
Click the command to copy.
Twenty minutes. If you need SSO, self-hosting, or a DPA, we start here, and end with a signature.
Book 20 minutes → Typical reply: same working day, CET.