CPersona

Recall fusion mode (CPERSONA_RECALL_MODE)

- rrf(default) — Reciprocal Rank Fusion: merges the vector + FTS channels by rank only. Robust and scale-free, but discards score magnitude.
- rsf— Relative Score Fusion: per-query min-max-normalizes each channel's raw score (cosine for vector, bm25 for keyword) and sums them, so the keyword channel's bm25 magnitude survives the merge.Recommended for topic-drift-prone or space-less language (e.g. Japanese) contexts, where that magnitude is the discriminating signalrrfflattens away (≈ Weaviate'srelativeScoreFusion; see the ClotoCoreRECALL_CONTAMINATION_AB_2026-06-14report §10–12).Caveat:min-max normalization can over-cut small, closely-scored result sets whenautocutis enabled —rrfremains the default until that interaction is hardened.
- cascade— Sequential channel fill (legacy).

WithCPERSONA_CONFIDENCE_ENABLED=true, the fusion mode does not decide the order you get back.Fusion selects which candidates enter the result set; confidence scoring then re-sorts that set, and the quality gate keys on the confidence score rather than on the fused one. Measured on a 1,545-document corpus with 394 queries: with confidence on,rsfandrrfreturned the same rows in the same order forall 394queries; with it off, the two agreed on fewer than 10%. So if you set a fusion mode expecting a ranking change, either leave confidence off, or expect the mode to affect which memories are considered and not the order they come back in.

- ~10,700 LOCPython across focused modules, plus a 3,300-line vendored MCP common snapshot
- 819 test functionsacross 73 test modules — 1,030 cases once the behavioural matrix is parametrised (~24,100 LOC, more test code than server code), including structural-enforcement gates
- Schema v13(auto-migrating)
- MIT License

cpersona is an MCP server — it works with any MCP-compatible host:

- ](https://github.com/Cloto-dev/cpersona/blob/master/benchmarks/README.md)Claude Desktop
- Claude Code
- ClotoCore(AI agent platform, where cpersona originated)
- Any custom MCP client

cpersona is the memory layer ofClotoCore, an open-source AI agent platform written in Rust. While cpersona is fully standalone (MIT license), it was designed to give AI agents persistent, searchable memory within the ClotoCore ecosystem.

- Audit-gated releases— before a release is cut, the codebase goes through comprehensive multi-agent audit rounds (independent finders per dimension, each finding adversarially verified from multiple lenses). v2.4.39 shipped after three such rounds — 43 fixes, every one re-verified against the tree it landed on.
- Issue registry— every audited defect lives inqa/issue-registry.jsonwith a machine-checkable code pattern;scripts/verify-issues.shverifies that every fix marker is still present (and every removed defect stays removed), so a regression or a silently-reverted fix fails loudly.
- Structural CI gates— invariants that a plain test can't express are enforced by AST- and behaviour-level gates in the pytest suite (run in CI on Python 3.11/3.13): every writer holds the shared write lock, agent-scoped SQL carries its isolation predicates, identity/dedup probes carry the project/channel axes, andcheck_healthperforms no embedding network I/O while holding the write lock.
- Release lifecycle standard— the release process itself is specified indocs/RELEASE_LIFECYCLE_STANDARD.md(v1.0), piloted in this repository as the reference implementation for Cloto-family projects.

Known issue:v2.4.39 and earlier under-scan vector recall on corpora beyond a few hundred rows (bug-085; v2.4.38–v2.4.39 are the most affected — the limit clamp closed the only workaround). Fixed in v2.4.40; upgrading is strongly recommended. SeeSUPPORT.md § Known issues.

Full policy:SUPPORT.md· specification:RELEASE_LIFECYCLE_STANDARD.md· security reports:[SECURITY.md.