CogniCore

I'm a solo dev. Six months ago I started CogniCore (https://github.com/cognicore-dev/cognicore-env) — an open-source memory layer for AI agents. The problem it solves: your coding agent relearns your repo every single session. Which build command works, which test is flaky, which import order breaks things — all gone the moment the context window closes. CogniCore records what the agent tried, what failed, what worked, with the evidence attached — and recalls it on the next run. Pure Python stdlib, SQLite storage, zero core dependencies. No vector DB, no embedding server. MIT. Where it is after six months: • ~13,000 downloads on PyPI (https://pypistats.org/packages/cognicore-env) — for transparency, that total includes CI/mirror pulls; the organic fraction is humbler. Both numbers are public, and I'd rather quote the big one with its caveat attached than let someone find it in a comment. • 69 stars, 12 forks, ~760 passing tests • MCP server included: pip install "cognicore-env[mcp]" then cognicore mcp serve — works with Claude, Cline, Cowork, anything MCP-compatible The part I'm actually proud of, from the last two weeks: • We proposed a portable memory-transfer format to https://github.com/mem0ai/mem0 (66k stars, the biggest open-source agent-memory project) and shipped the reference implementation as an upstream PR — in review now • A contributor stress-tested it on Windows, unprompted, and left a public verification report on the PR • Another landed a float-serialization security fix in our repo this week after we accepted his review finding — merged with 34/34 tests green • A third opened a cross-engine conformance PR for the canonicalization the format needs Nobody was asked, paid, or coordinated. They showed up because the design arguments happen in public issue threads and the claims stay boring. What it does beyond memory: a reflection engine that surfaces repeated failure patterns instead of just storing them, threat detection that blocks prompt injection at the agent layer, and Ed25519-signed memory transfer between agents — fail-closed, no valid signature means no import, and there is no flag to disable that. Honest limits, stated up front: 10 MCP tools on the released surface (we publicly corrected a "30 tools" overclaim when a contributor caught it — that thread did more for the project than any feature), TF-IDF/BM25 retrieval rather than embeddings by design, and signature enforcement currently lives in the integration layer — server-side enforcement across the whole MCP surface is the next release. If you build agents, or you've thought about what memory should be allowed to carry between two systems that don't trust each other — the repo (https://github.com/cognicore-dev/cognicore-env). The design threads in the issues are genuinely better reading than the README.