CPersona

MCP Memory Server

Give Claude persistent memory across sessions. Single SQLite file. 29 tools. Zero LLM dependency.

Quick Start · Features · Architecture · All Tools · PyPI · Zenn Book (JP)

Standalone repository — This is the standalone version for use with Claude Desktop, Claude Code, and any MCP client. If you are a ClotoCore user, install CPersona from the in-app marketplace (ClotoHub) instead — it distributes this same repository.

Project status (July 2026) — The 2.4 series is the Stable line (latest: v2.4.40, gated by three comprehensive audit rounds — see Quality Assurance). The 2.5 series is the Current line (latest: v2.5.3) — an internal stabilization line that has passed the full release gate and is where all fixes land, pending production-soak certification; the DB schema is preserved across the line, and feature development resumes in 2.6. v2.5.3 refuses to start the HTTP transport when CPERSONA_AUTH_TOKEN is unset, wherever it binds — earlier versions allowed an unauthenticated loopback bind, which a tunnel or reverse proxy silently turns into public exposure (bug-198, HIGH; see Remote HTTP transport). A deployment that deliberately runs without authentication must now say so with CPERSONA_ALLOW_UNAUTHENTICATED_HTTP=true; the stdio transport is unaffected. v2.5.2 changes three things about MCP tool responses — store reports its outcome in result (stored / skipped / rejected) instead of an always-true ok, check_health reports the single status verdict without the healthy boolean, and every tool-level failure a handler returns now carries ok: false (most used to return error alone, with no ok to branch on; the explanation still travels in error, except on store, which puts it in reason). Two shapes stay outside that rule and always did: the outermost MCP dispatch answers an unknown tool name, or an exception escaping a handler, with a bare error and no ok — that layer is vendored from a library shared with the other Cloto servers, so correcting it is an upstream change rather than a local edit — and a successful read (get_contents, list_memories, list_episodes, get_profile) returns its payload with no ok either. Branch on ok is false, and treat a response carrying error as a failure whether or not ok is present. Tiers and support windows: Release Channels & Support.

The Problem

Claude forgets everything between sessions. Every conversation starts from zero — no context about your project, your preferences, or what you discussed yesterday.

cpersona fixes this. It's an MCP server that stores memories in a local SQLite file and retrieves them through hybrid search. Claude remembers you.

Related MCP server: mcp-memory-graph

Quick Start

Prerequisites: Python 3.11+ (and uv for the one-command path).

Claude Code? Let the agent do the setup. This repo ships an Agent Skill that walks Claude through the whole installation — cpersona, the embedding server, MCP registration, and a store/recall smoke test — and, more importantly, teaches it when to store, recall, and archive memories afterwards:

# Installed from PyPI? The skill ships inside the wheel — no clone needed:
python -c "import cpersona,pathlib,shutil; s=pathlib.Path(cpersona.__file__).parent/'skills'/'cpersona-memory'; shutil.copytree(s, pathlib.Path.home()/'.claude/skills/cpersona-memory', dirs_exist_ok=True)"

# Running via uvx (isolated environment), or not installed yet:
git clone --depth 1 https://github.com/Cloto-dev/cpersona.git /tmp/cpersona
mkdir -p ~/.claude/skills && cp -r /tmp/cpersona/skills/cpersona-memory ~/.claude/skills/

Then just tell Claude Code: "Set up CPersona — I want persistent memory." The manual steps below are for Claude Desktop users and anyone who prefers to configure things by hand.

1. Install cpersona

uvx cpersona          # run directly, no install step
# or
pip install cpersona  # then the `cpersona` command is on your PATH

git clone https://github.com/Cloto-dev/cpersona.git
cd cpersona
python -m venv .venv
source .venv/bin/activate      # Windows: .venv\Scripts\activate
pip install .

Run it with python -m cpersona (or python server.py).

2. Set up Embedding Server (Recommended)

cpersona's hybrid search works best with an embedding server for vector similarity. cpersona is embedding-server-agnostic: point CPERSONA_EMBEDDING_URL (see step 3) at any HTTP endpoint that implements the following minimal contract.

POST /embed
Request:  { "texts": ["string", ...] }        # non-empty array, max 100 per batch
Response: { "embeddings": [[float, ...], ...], "dimensions": <int> }

Contract requirements (2.5.0b1 clarifications):

• Embeddings MUST be L2-normalized. cpersona computes similarity as a raw dot product; a backend returning unnormalized vectors biases ranking by vector magnitude. Every supported backend (the client's api mode and all CEmbedding providers) already normalizes.

• The contract is role-less — queries and documents are embedded through the same call. Prompt-prefix models (e5-style, prompted bge) will underperform behind it; symmetric or retrieval-merged models (jina-v5-nano, bge-m3, MiniLM) are the intended fit.

• Swapping models behind the same URL: cpersona fingerprints the backend by embedding dimension only (the contract carries no model identity). A same-dimension model swap silently invalidates the stored corpus — after one, re-embed (check_health(fix=true) repairs NULLed rows) and calibrate_threshold.

The reference server is CEmbedding (MIT) — it runs jina-v5-nano on-device (CPU) and exposes exactly this endpoint:

git clone https://github.com/Cloto-dev/CEmbedding.git && cd CEmbedding
python -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install ".[onnx]"
python download_model.py --model jina-v5-nano
EMBEDDING_PROVIDER=onnx_jina_v5_nano python server.py   # serves http://127.0.0.1:8401/embed

cpersona ships with defaults tuned against jina-v5-nano (768d). Any other server that satisfies the contract above works too — see Benchmarks for models with published measurements.

Without an embedding server, cpersona falls back to FTS5 + keyword search only. Vector search (the strongest retrieval layer) will be disabled.

3. Configure your MCP client

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "cpersona": {
      "command": "uvx",
      "args": ["cpersona"],
      "env": {
        "CPERSONA_DB_PATH": "/home/you/.claude/cpersona.db",
        "EMBEDDING_MODE": "http",
        "EMBEDDING_HTTP_URL": "http://127.0.0.1:8401/embed"
      }
    }
  }
}

The embedding server from step 2 is a plain HTTP process, not an MCP server — run it however you run background services (a terminal, launchd/systemd, etc.); cpersona only needs its URL.

Windows: use C:/Users/you/.claude/cpersona.db for the DB path. No embedding server yet? Drop the two EMBEDDING_* lines (or set EMBEDDING_MODE=none) — cpersona runs on FTS5 + keyword and tells you when it's degraded.

Claude Code:

claude mcp add-json cpersona '{"type":"stdio","command":"uvx","args":["cpersona"],"env":{"CPERSONA_DB_PATH":"/home/you/.claude/cpersona.db","EMBEDDING_MODE":"http","EMBEDDING_HTTP_URL":"http://127.0.0.1:8401/embed"}}' -s user

That's it. Claude now has persistent memory. Ask it to store something and recall it in a later session.

Features

Hybrid Search — Three independent retrieval strategies run in parallel and merge results via Reciprocal Rank Fusion (RRF):

Memory Types:

• Declarative memory — Individual facts, decisions, instructions stored via store

• Episodic memory — Conversation summaries archived via archive_episode

• Profile memory — Accumulated user/project attributes via update_profile

Confidence Scoring — Each recalled memory gets a confidence score combining:

• Cosine similarity (semantic relevance)

• Dynamic time decay (adapts to corpus time range — a 1-year-old corpus and a 1-day-old corpus use different decay curves)

• Recall boost (frequently useful memories surface more easily, with natural fade-out)

• Completion factor (resolved topics decay faster)

Zero LLM Dependency — cpersona is a pure data server. It never calls an LLM internally. All summarization and extraction is performed by the calling agent. This means zero API costs from cpersona itself, deterministic behavior, and no hidden latency.

Additional capabilities:

• Agent namespace isolation — multiple agents share one DB without interference

• Background task queue — DB-persisted, crash-recoverable async processing

• JSONL export/import — full memory portability between environments

• Agent-to-agent memory merge — atomic copy/move with deduplication

• Auto-calibration — statistical threshold tuning via null distribution z-score (no labels needed)

• Health check — a check registry with severity-tagged issues (critical/warn/info) and auto-repair (contamination, duplicates, FTS integrity, embedding dimension drift, schema objects, isolation-axis hygiene, stale tasks, invalid data), plus a python -m cpersona.checkup CLI for CI gating

• Deep check — semantic data quality analysis (anonymous source recovery, short content, stale profiles, orphaned episodes)

• Memory protection — lock/unlock to prevent accidental deletion or editing

• Recent recall penalty — suppresses echo chamber effect for frequently recalled memories

• stdio + Streamable HTTP transport

• Single-file SQLite — no external database required

Architecture

                         ┌─────────────────────────────────────┐
                         │            MCP Host                 │
                         │   (Claude Desktop / Claude Code)    │
                         └──────────────┬──────────────────────┘
                                        │ MCP (JSON-RPC)
                         ┌──────────────▼──────────────────────┐
                         │           cpersona                  │
                         │         (server.py)                 │
                         │                                     │
                         │  ┌─────────┐  ┌─────────┐          │
                         │  │  store   │  │ recall  │  ...     │
                         │  └────┬────┘  └────┬────┘          │
                         │       │             │               │
                         │  ┌────▼─────────────▼────────────┐  │
                         │  │         SQLite DB              │  │
                         │  │                                │  │
                         │  │  memories    (content + embed) │  │
                         │  │  episodes    (summaries)       │  │
                         │  │  profiles    (attributes)      │  │
                         │  │  memories_fts (FTS5 index)     │  │
                         │  │  episodes_fts (FTS5 index)     │  │
                         │  │  pending_memory_tasks (queue)  │  │
                         │  └────────────────────────────────┘  │
                         │                                      │
                         └──────────────┬───────────────────────┘
                                        │ HTTP
                         ┌──────────────▼──────────────────────┐
                         │       Embedding Server              │
                         │  (jina-v5-nano ONNX, 768d)          │
                         └─────────────────────────────────────┘

Recall flow (RRF mode):

Query → ┌── Vector search (cosine similarity)  ──┐
        ├── FTS5 search (episodes + memories)    ──┼── RRF merge → Confidence scoring → Top-K
        └── Keyword fallback                     ──┘

Benchmarks

Measured on LMEB (Long-horizon Memory Embedding Benchmark, arXiv:2603.12572) — 22 datasets subsuming LoCoMo and LongMemEval, measured here as 22 retrieval tasks. The metric is Mean NDCG@10 across all 22 tasks.

Two tracks isolate the pipeline's contribution:

• Track A — the raw embedding model alone (baseline retrieval).

• Track B — the same embeddings routed through cpersona's real store/recall code paths: SQLite + FTS5 + RRF fusion + per-agent auto-calibration (cpersona v2.4.40, full-ranking regime).

cpersona's hybrid pipeline outranks the raw embedding on both models (Track B > Track A) — the fusion layers add signal rather than merely persisting vectors. The weaker the embedding, the larger the pipeline's contribution: the FTS5/keyword layers rescue queries the vector search alone misses. Methodology, the measurement harness, and the reproduction regime live in benchmarks/.

All Tools

Configuration

All settings via environment variables with sensible defaults:

The generic aliases EMBEDDING_MODE / EMBEDDING_HTTP_URL / EMBEDDING_MODEL are also accepted (the CPERSONA_-prefixed form wins when both are set) — the marketplace catalog and the Quick Start use the generic names.

Remote (HTTP) transport

The default transport is stdio, where the MCP client owns the process and no network is involved. Set CPERSONA_TRANSPORT=streamable-http to serve over HTTP instead — one server, several clients, reachable over a network.

A loopback bind is not a security boundary. Tunnels (cloudflared, ngrok), reverse proxies, kubectl port-forward and published container ports all forward to 127.0.0.1, so binding there says nothing about who can reach the port. Every tool is exposed to whoever can — including delete_agent_data and the file-reading/writing export_memories / import_memories. Set CPERSONA_AUTH_TOKEN whenever the process is not something only you can talk to.

Since v2.5.3 the server enforces that: with CPERSONA_TRANSPORT=streamable-http and no CPERSONA_AUTH_TOKEN, it refuses to start. If you are upgrading from 2.5.2 or earlier and run the HTTP transport without a token, it will not start — set CPERSONA_AUTH_TOKEN, or set CPERSONA_ALLOW_UNAUTHENTICATED_HTTP=true to state that you really do want no authentication (local development only). Earlier versions allowed an unauthenticated loopback bind and logged that it was "bound to loopback only", which read as an all-clear and was not one.

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 signal rrf flattens away (≈ Weaviate's relativeScoreFusion; see the ClotoCore RECALL_CONTAMINATION_AB_2026-06-14 report §10–12). Caveat: min-max normalization can over-cut small, closely-scored result sets when autocut is enabled — rrf remains the default until that interaction is hardened.

• cascade — Sequential channel fill (legacy).

Stats

• ~10,700 LOC Python across focused modules, plus a 3,300-line vendored MCP common snapshot

• 608 test functions across 54 test modules — 710 cases once the behavioural matrix is parametrised (~17,800 LOC, more test code than server code), including structural-enforcement gates

• Schema v13 (auto-migrating)

• MIT License

Works With

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

• Claude Desktop

• Claude Code

• ClotoCore (AI agent platform, where cpersona originated)

• Any custom MCP client

Part of ClotoCore

cpersona is the memory layer of ClotoCore, 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.

Quality Assurance

Every release is gated by a machine-verifiable quality process:

• 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 in qa/issue-registry.json with a machine-checkable code pattern; scripts/verify-issues.sh verifies 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, and check_health performs no embedding network I/O while holding the write lock.

• Release lifecycle standard — the release process itself is specified in docs/RELEASE_LIFECYCLE_STANDARD.md (v1.0), piloted in this repository as the reference implementation for Cloto-family projects.

Release Channels & Support

Releases follow a three-tier model — Stable (production-certified, critical fixes only), Current (newest release line, all fixes land here), and Experimental (alpha/beta pre-releases, opt-in). When a new line is certified Stable, the previous one keeps critical-fix support for 30 more days, then reaches EOL. Current status: 2.4.x is the Stable line (latest v2.4.40) and 2.5.x is the Current line (latest v2.5.3), where all fixes land while it awaits production-soak certification.

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. See SUPPORT.md § Known issues.

Known issue (Stable line): the 2.4.x line starts the HTTP transport unauthenticated when CPERSONA_AUTH_TOKEN is unset and the bind is a loopback address (bug-198, HIGH). A loopback bind does not bound reachability — tunnels, reverse proxies, kubectl port-forward and published container ports all forward to 127.0.0.1. If you serve 2.4.x over HTTP, set CPERSONA_AUTH_TOKEN. The startup enforcement that makes a missing token refuse to boot wherever it binds is in the Current line (v2.5.3); the Stable line only warns.

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

Found a bug, or something the docs do not explain?

Open an issue — bug report or feature request.

Reports are welcome even when you are not certain it is a bug. If it turns out to be a configuration problem, that is still useful signal — it means the documentation was unclear, which is a defect of its own. Security vulnerabilities are the one exception: please report those privately via SECURITY.md rather than in a public issue.

Learn More

• Zenn Book (Japanese) — Full design walkthrough and setup guide

• Replacing /compact with external memory (Japanese) — Measured token economics of the session-end → /clear → recall workflow

• Memory System Design — Technical specification

• ClotoCore — The AI agent platform

License

MIT — free to use from any MCP host without restriction.