ccRecall

中文版

A local memory service for Claude Code — indexes your conversation history, recalls relevant context on demand, and injects it into future sessions. Zero API cost for the core service; optional post-session extraction via Haiku adds ~$0.001/session.

Manifesto

Five stances, non-negotiable:

1. Local-first. Your machine, your data. No cloud, no accounts, no services beyond localhost.
2. SQLite is the API. One .db file. sqlite3 to inspect anything. Your memories are never a black box.
3. FTS5, not vectors. Code conversations search for file paths, error messages, tool names — keywords, not vibes. At hundreds of sessions, embeddings solve a problem that doesn't exist.
4. LLM distills, human curates. Haiku extracts post-session (~$0.001). Auto memory holds decisions you name. ccRecall holds the long tail no one writes down by hand.
5. Read-only toward ~/.claude/. Reads your JSONL logs, never writes to Claude Code's state. Worst case: bad search results. Never: a broken setup.

Core Concept

Every time you start a new Claude Code session, the AI forgets everything. The architecture you spent 20 minutes explaining, the bug you debugged together, the decisions you made — all gone. You start over.

CLAUDE.md and RESUME.md help, but they're static files you maintain by hand. ccRecall automates this: it reads your JSONL conversation logs, builds a searchable index, and serves relevant memories back to Claude Code through hooks and MCP tools. The AI remembers what it learned — you don't have to remind it.

ccRecall is the "memory" counterpart to ccRewind (a conversation replay GUI). ccRewind lets humans look back at what happened; ccRecall lets the AI remember what happened.

Note: This project is unrelated to spences10/ccrecall, an analytics-focused tool that happens to share the name. Because the npm package ccrecall is already taken, we publish as @tznthou/ccrecall and the CLI binary is named ccmem.

Dogfood baseline

Single-user daily-driver data, not a controlled study. Updated 2026-09-11.

| Metric | Value | |--------|-------| | Running continuously | 147 days | | Sessions indexed | 1,538 main sessions across 106 projects | | Memories stored | 1,639 (98% keyed for dedup) | | Topics in knowledge map | 35,887 | | DB on disk | 102 MB |

Counted the way /health reports them: mainSessionCount excludes subagent sessions, which are indexed separately (1,648 more). Earlier revisions of this table did not say which count they used, so the session figure is not comparable across versions.

n = 1. These numbers show the system runs and accumulates data — not that every memory is useful. We don't yet have a good metric for "did this memory actually help the session." That's an open problem.

What the numbers don't show: without ccRecall, every session starts cold — you re-explain context the AI already learned yesterday. With it, startup injection surfaces the relevant memory and the session picks up mid-thought. One injection that saves a five-minute preamble pays for the entire daemon.

Features

| Feature | Description | |---------|-------------| | Rule-based summarization | Extracts intent, activity, outcome, and tags from sessions — no LLM calls, zero API cost | | FTS5 full-text search | Sub-100ms keyword search across all conversation history, fast enough for hook injection | | CJK / mixed-script search | Trigram tokenizer indexes Chinese / Japanese / Korean text; per-token AND LIKE fallback handles short queries like UI 記憶 that trigram alone can't match | | Incremental indexing | Only re-indexes sessions that changed (mtime diffing), handles resumed sessions via UUID dedup | | Metacognition | knowledge_map aggregates topic mentions from sessions + memories. Depth derived from mention count (shallow / medium / deep). Exposed via MCP recall_context | | Forgetting curve | Memories compress over time: raw → summary → one-liner → deleted. Confidence decays on unused memories. Background maintenance tick runs every 5 min | | Cross-project memory (v0.4.1) | Memories surface across projects via topic intersection — if two projects share a knowledge_map topic, high-confidence memories from one appear in the other's startup injection (max 3 rows, confidence ≥ 0.8 gate) | | Mid-conversation recall | SessionStart fires once; the UserPromptSubmit hook is the second trigger point. Topics from the prompt drive an inverse-document-frequency ranked lookup, suppressed per session and capped at 8 memories. 300 ms timeout, fails silent | | Watch mode | chokidar-based JSONL watcher picks up new sessions within 2 s; periodic 10 min full-resync covers missed filesystem events | | Rescue reindex | /session/end and /session/last both retry a reindex on miss, and /session/last staleness-gates via notBefore — no fresh-session race between the hook, the wrapper, and the daemon | | Auto-start (macOS) | ccmem install-daemon registers a LaunchAgent so the service stays up across reboots | | Read-only | Never modifies ~/.claude/ — only reads JSONL logs |

Architecture

flowchart TB
    subgraph Input["Data Source (read-only)"]
        JSONL["~/.claude/projects/*/*.jsonl"]
    end

    subgraph Core["ccRecall Service (port 7749)"]
        Watcher["Watcher<br/>chokidar, 2 s debounce"]
        Scanner["Scanner<br/>find JSONL files"]
        Parser["Parser<br/>parse conversations"]
        Summarizer["Summarizer<br/>rule-based extraction"]
        DB[("SQLite + FTS5<br/>index & search")]
        API["HTTP API<br/>6 endpoints"]
    end

    subgraph Consumers["Consumers"]
        Hook["Claude Code Hooks<br/>SessionStart / UserPromptSubmit / SessionEnd"]
        Wrapper["Extraction wrapper<br/>post-session Haiku pass"]
        MCP["MCP Server<br/>recall_query / recall_save"]
    end

    JSONL --> Watcher --> Scanner --> Parser --> Summarizer --> DB
    DB <--> API
    Hook -->|"inject / end-confirm"| API
    Wrapper -->|"GET /session/last"| API
    MCP <-->|"shared SQLite via WAL"| DB

Arrows point in the direction of the call: hooks and the extraction wrapper talk to the daemon over HTTP, while the MCP server is a separate process that opens the same SQLite file directly (WAL mode) — it never goes through the HTTP API.

Session lifecycle

How the three injection points and the two timing-sensitive ends actually flow:

sequenceDiagram
    participant CC as Claude Code
    participant H as Hooks
    participant R as ccRecall
    participant W as Wrapper
    participant X as Haiku extractor

    rect rgb(235, 244, 255)
    Note over CC,R: Session start
    CC->>H: SessionStart
    H->>R: GET /memory/startup
    R-->>H: 3-tier memory pick (< 400 tokens)
    H-->>CC: inject into context
    end

    rect rgb(240, 240, 235)
    Note over CC,R: Every prompt (L1)
    CC->>H: UserPromptSubmit
    H->>R: GET /memory/prompt (300 ms timeout)
    R-->>H: topic-matched rows, deduped per session
    H-->>CC: inject, or stay silent
    end

    Note over CC,R: Meanwhile the watcher<br/>reindexes JSONL writes (2 s debounce)

    rect rgb(255, 244, 235)
    Note over CC,X: Session end
    CC->>H: SessionEnd
    H->>R: POST /session/end (index confirm + rescue)
    W->>R: GET /session/last?notBefore=launch_ts
    R-->>W: sessionId (staleness-gated, subagent-filtered)
    W->>X: text-only transcript
    X->>R: recall_save × 0–5 (via MCP)
    end

The notBefore gate keeps a stale session from being extracted twice; the subagent filter keeps an Agent-tool side-session from shadowing the real one. Both guards exist because the wrapper queries within seconds of session close, racing the watcher.

Tech Stack

| Technology | Purpose | Notes | |------------|---------|-------| | Node.js 20–22 + TypeScript | Runtime | ES modules, strict mode | | better-sqlite3 | Database | Synchronous API, zero external deps | | FTS5 | Full-text search | Built into SQLite, trigram tokenizer with LIKE fallback for short CJK / mixed-script queries | | Native http | HTTP server | No Express — minimal surface, localhost only | | chokidar | Filesystem watcher | Cross-platform JSONL change detection with 2 s debounce + single-flight | | vitest | Testing | 651 tests across 40 files, integration-style | | @modelcontextprotocol/sdk | MCP server | stdio transport, shared SQLite via WAL |

Quick Start

First time here? The full walkthrough (install via npm → MCP setup → everyday usage) lives in docs/tutorial.md. The section below is the contributor / dev-mode path.

Prerequisites

• Node.js >=20.0.0,<23.0.0
• pnpm

Installation

git clone https://github.com/tznthou/ccRecall.git
cd ccRecall

pnpm install

# Start development server (auto-indexes on startup, watches ~/.claude/projects)
pnpm dev

The service starts at http://127.0.0.1:7749 and indexes all JSONL files in ~/.claude/projects/.

Verify

# Health check — should show mainSessionCount > 0
curl http://127.0.0.1:7749/health

# Search your conversation history
curl "http://127.0.0.1:7749/memory/query?q=authentication&limit=5"

API Endpoints

Six endpoints, each with a live caller — v0.5.0 removed the other eight (/journal/*, /memory/save, /memory/context, /metacognition/check, /session/checkpoint, /lint/warnings); they now return 404.

| Endpoint | Method | Description | Caller | |----------|--------|-------------|--------| | /health | GET | Service health + DB stats + integrity check status | CLI, extraction wrapper | | /memory/startup?project=... | GET | SessionStart-tier retrieval: cold + recent-confidence + FTS fallback, token-budgeted | SessionStart hook | | /memory/query?q=...&limit=...&project=... | GET | FTS5 search across memories with optional project filter | SessionStart hook (keyword tier) | | /memory/prompt?project=...&q=... | GET | Mid-conversation recall: topic lookup against the current prompt, per-session dedup and ceiling | UserPromptSubmit hook | | /session/end | POST | Confirm the just-ended session is indexed (rescue reindex on miss) | SessionEnd hook | | /session/last?cwd=... | GET | Most recent session metadata for a project path (staleness-gated via notBefore) | Extraction wrapper |

MCP Tools

| Tool | Purpose | |------|---------| | recall_query | User-scoped FTS5 keyword search across memories with project-aware ranking. Cross-project memories surface via topic intersection | | recall_context | Topic-clustered retrieval — normalizes keywords, groups memories by matched topic with depth signals, falls back to per-keyword FTS if no topic matches | | recall_save | Store a new memory. An optional key slug dedups (same key updates instead of duplicating), and topics are auto-extracted for cross-project retrieval. Optional origin (explicit | agent-inferred) says who is writing: an agent-inferred save is refused against an existing explicit memory, so extraction cannot overwrite what you wrote by hand. Optional messageId records which message the memory came from — checked against that session's real messages, and dropped rather than stored if it doesn't match |

Memory types (for recall_save):

• decision — explicit choice with rationale
• discovery — non-obvious finding
• preference — user style or convention
• pattern — recurring workflow or code template
• feedback — user correction on past work

Expose them to Claude Code. After pnpm build, the ccmem-mcp bin is on the repo's node_modules/.bin path — point claude mcp add at it or at a global install:

# Using the built bin (after pnpm build)
claude mcp add ccrecall --scope user -- /absolute/path/to/ccRecall/dist/mcp/server.js

# Or using tsx for development (no build step)
claude mcp add ccrecall --scope user -- /absolute/path/to/ccRecall/node_modules/.bin/tsx /absolute/path/to/ccRecall/src/mcp/server.ts

A ready-to-copy example lives at .mcp.json.example.

See hooks/README.md for SessionStart / UserPromptSubmit / SessionEnd hook installation.

CLI Commands

@tznthou/ccrecall ships two binaries:

• ccmem — daemon launcher + admin commands
• ccmem-mcp — MCP server (registered with Claude Code via claude mcp add)

Daemon and hook lifecycle (macOS):

| Command | Purpose | |---------|---------| | ccmem | Run the daemon in foreground | | ccmem install-daemon | Register a LaunchAgent (auto-start at login) | | ccmem uninstall-daemon | Stop and remove the LaunchAgent | | ccmem install-hooks | Merge SessionStart / SessionEnd / UserPromptSubmit entries into ~/.claude/settings.json | | ccmem uninstall-hooks | Remove ccRecall's hook entries (other hooks untouched) | | ccmem cleanup --orphans | List (and with --yes, delete) memories whose session rows are gone |

ccmem promote / ccmem reject were removed in v0.5.0 along with the journal pipeline they reviewed.

ccRecall vs auto memory

ccRecall lives alongside Claude Code's built-in auto memory (~/.claude/projects/*/memory/). They're complementary — use them for different things.

| | auto memory | ccRecall | |---|---|---| | Write path | Claude curates by hand — new .md file + MEMORY.md index line | Two tracks: rule-based session summaries (automatic, free) + opt-in Haiku extraction (~$0.001/session, where most searchable memories come from) | | Read path | Always in session context (MEMORY.md loads at session start) | On-demand MCP query when auto memory has no entry | | Signal density | High — facts worth naming | Long tail — everything the hook can extract | | Typical use | "Remember X" / "always Y" — durable preferences, decisions | "Didn't we fix that?" / "last time" — cross-session recall |

Default for saving: write to auto memory, let the hook harvest ccRecall independently. Don't call recall_save to mirror a fact you already curated — duplicate writes just create noise.

Default for querying: MEMORY.md is already in context — check the index first. Fall back to recall_query / recall_context only when the user references past work and auto memory has no matching entry.

ccRecall's value is the long tail that auto memory can't cover (nobody hand-curates 500 sessions of notes). If Claude defaults to both, auto memory wins because it's already loaded and curated. ccRecall earns its keep when the curated index misses.

What if Anthropic builds this in? Auto memory today is .md files tied to Claude Code's project structure — portable as text, but not queryable. ccRecall is a single SQLite file you can inspect with sqlite3, query with SQL, back up by copying one file, and use independently of Claude Code's config. The difference isn't local vs. cloud — it's a curated document store vs. a searchable database. If built-in memory covers your needs, use it. ccRecall is for when you want structured recall across hundreds of sessions that no one will curate by hand.

Running as a service (macOS)

ccRecall runs as a local HTTP daemon. To keep it up across reboots, register a per-user LaunchAgent:

pnpm build
node dist/index.js install-daemon        # or `ccmem install-daemon` if globally linked
node dist/index.js install-daemon --dry-run   # preview plist without writing

# verify
launchctl list | grep ccrecall
curl http://127.0.0.1:7749/health

# remove
node dist/index.js uninstall-daemon

The installer:

• writes ~/Library/LaunchAgents/com.tznthou.ccrecall.plist
• routes logs to ~/Library/Logs/ccrecall/ccrecall.{out,err}.log
• propagates CCRECALL_PORT / CCRECALL_DB_PATH from the current shell into the plist, so the LaunchAgent uses the same settings as your interactive run
• refuses to touch a plist whose Label isn't ccRecall's (safety check)

Full manual-install, troubleshooting, and uninstall docs: docs/launchd.md.

Linux/Windows equivalents (systemd unit, Windows service) are planned for a future release. For now, run under nohup or your process manager of choice.

Monitoring

The daemon runs PRAGMA integrity_check on startup and every 6 hours. The result (timestamp + boolean) is cached and surfaced on /health as lastIntegrityCheckAt / lastIntegrityCheckOk. When drift is detected, the full integrity_check output is written to a timestamped file under ~/.ccrecall/integrity-alerts/.

If you see a drift alert, snapshot the DB before running REINDEX. REINDEX fixes the symptom but destroys the forensic state:

cp ~/.ccrecall/ccrecall.db ~/ccrecall-drift-snapshot.db
sqlite3 ~/.ccrecall/ccrecall.db 'REINDEX;'

WAL maintenance

Each indexer batch ends with PRAGMA wal_checkpoint(TRUNCATE) so the ccrecall.db-wal sidecar is reset to 0 bytes after every reindex. On a long-running daemon you should see WAL hovering near 0 most of the time, spiking briefly while a batch runs.

If you ever see WAL growing unboundedly (close to the size of the main DB), check stderr for [indexer] WAL checkpoint busy warnings — that means a reader has been holding a snapshot past busy_timeout across several consecutive batches and the truncate keeps deferring. Identify the offending client and the next clean batch will reclaim the space.

Project Structure

ccRecall/
├── src/
│   ├── core/
│   │   ├── types.ts              # All type definitions
│   │   ├── parser.ts             # JSONL conversation parser
│   │   ├── scanner.ts            # File system scanner
│   │   ├── summarizer.ts         # Rule-based session summarizer
│   │   ├── topic-extractor.ts    # Rule-based topic extraction
│   │   ├── database.ts           # SQLite + FTS5 (trimmed from ccRewind)
│   │   ├── indexer.ts            # Indexing pipeline orchestrator
│   │   ├── memory-service.ts     # Memory lifecycle (touch / delete / update)
│   │   ├── compression.ts        # L0→L1→L2→delete state machine
│   │   ├── maintenance-coordinator.ts  # Background compression tick
│   │   ├── watcher.ts            # chokidar JSONL watcher (Phase 4e)
│   │   └── log-safe.ts           # scrubErrorMessage — log-injection defence
│   ├── api/
│   │   ├── server.ts             # HTTP server
│   │   └── routes.ts             # Request routing + rescue reindex
│   ├── mcp/
│   │   ├── server.ts             # MCP stdio server entry (shebang bin)
│   │   └── tools.ts              # recall_query + recall_context + recall_save
│   ├── cli/
│   │   └── daemon.ts             # install-daemon / uninstall-daemon (macOS)
│   └── index.ts                  # HTTP entry point + subcommand dispatch
├── hooks/
│   ├── session-start.mjs         # Inject memories on SessionStart (stdout)
│   ├── session-end.mjs           # POST /session/end on SessionEnd
│   └── README.md                 # Hook installation guide
├── docs/
│   ├── tutorial.md               # End-user walkthrough (install → MCP → usage)
│   ├── architecture.md           # Daemon design rationale (contributor-oriented)
│   └── launchd.md                # macOS LaunchAgent install/troubleshoot
├── tests/                        # 651 tests across 40 files (parser, scanner,
│   │                             # summarizer, database, indexer, e2e, MCP,
│   │                             # memories, hooks, watcher, CLI, migrations,
│   │                             # FTS5 CJK edge cases, integrity monitor, ...)
│   └── fixtures/                 # Sample JSONL + shared test helpers
├── .mcp.json.example             # MCP client config template
└── NOTICE / SECURITY.md / CONTRIBUTING.md / CODE_OF_CONDUCT.md

Related Projects

• ccRewind — Session replay GUI for Claude Code. ccRecall's core modules (parser, scanner, summarizer, database, indexer) were extracted from ccRewind.

Philosophy

Why this exists

Thariq from Anthropic's Claude Code team wrote about context management in April 2026 — 11,908 bookmarks, because everyone saved it to re-read but nobody had the tools to actually do it. He described the problem perfectly: context rot degrades model performance in long sessions, and autocompact fires at the worst possible moment.

He gave methodology, not tools. ccRecall is the tool.

The real trigger was simpler: I kept re-explaining the same architecture to Claude Code across sessions. Not because the AI is bad at remembering — it literally can't. Every session starts from zero. CLAUDE.md helps, but it's a static file I maintain by hand. The maintenance cost grows faster than the value. Sound familiar? That's exactly why humans abandon wikis too (Karpathy's LLM Wiki insight).

Design stances

Each flows from the manifesto. Each was a conscious choice against a popular alternative.

Rule-based, with a ceiling. The daemon uses heuristic extraction (regex patterns, tool usage analysis, outcome inference) for session summaries — zero API cost. For that job, "Edit x8, 5 files, committed" is more useful than a paragraph of prose. But rule-based has a ceiling: it handles structured signals (tool calls, file edits, commit messages) well; it misses discussion-only decisions, nuanced trade-offs, and anything that lives in natural language. That ceiling is why v0.4.1 added post-session Haiku extraction (~$0.001/session) — an LLM pass that catches what heuristics can't. In practice, the memories I actually search for mostly come from Haiku extraction or manual recall_save, not the rule-based layer. The daemon itself never calls an LLM; the LLM pass is opt-in and runs outside the daemon as a separate process.

FTS5, not vector search. Semantic search sounds better on paper, but for conversation logs — specific tools, file paths, error messages — keyword matching wins. FTS5 queries run in <10ms locally. No embedding model, no Chroma, no Docker container. At hundreds of sessions (not millions of documents), Karpathy's own analysis confirms: "plain index + keyword search is already sufficient under 500 sources."

HTTP + MCP dual interface. MCP tools are the most stable way to inject context into Claude (pull-based, Claude decides when to fetch). SessionStart hooks (push-based, automatic) are also stable. ccRecall runs both: HTTP for hooks, MCP for on-demand queries. Same SQLite backend, two access patterns.

Read-only, unconditionally. ccRecall never modifies ~/.claude/, never writes to session files, never injects itself into Claude Code's config. This isn't politeness — it's a trust boundary. If a background service can write to your config, one bug could corrupt your sessions. The user explicitly configures hooks and MCP. ccRecall doesn't install itself.

The excluded stack. No Docker — deployment friction for what should be a pnpm dev experience. No Electron — ccRecall has no UI (that's ccRewind's job). No vector database — solves a problem we don't have at this scale. These are deliberate exclusions, not missing features.

No opinionated injection. ccRecall doesn't decide what Claude should remember. It provides a search API — the injection layer presents results, Claude integrates them. Opinionated memory selection is a premature optimization that would be wrong in ways we can't predict.

Roadmap

| Version | Theme | Status | |---------|-------|--------| | v0.3.x | Manual save, automatic recall — memories come from explicit recall_save calls; SessionStart hook and MCP tools inject them into future sessions | Released | | v0.4.x | Post-session extraction via Haiku, cross-project memory via topic intersection, extraction pipeline hardening (race gates, compression integrity, subagent filtering, security) | Released | | v0.5.0 | The knife field: journal/scorer/harvester pipelines removed (zero promotions ever), endpoints 13 → 5, message_uuids dual-hash rebuild (DB 114MB → 42MB) | Released | | v0.5.2 | Recall weighting: log-compressed half-life decay, FTS relevance-first ranking via (-rank)*sqrt(EC) | Released | | v0.5.3 | CJK topic alignment: Han-aware topic extraction (\p{Script=Han} tokenizer + 32 Chinese stopwords + particle split), session-less rebuild fix | Released |

Tracked in GitHub Issues.

Changelog

Release notes and version history live in CHANGELOG.md. Every tagged version has a matching entry; the Unreleased section tracks what's landed on main but not yet published to npm.

License

Licensed under the Apache License, Version 2.0 — see LICENSE.

Copyright 2026 tznthou

Author

tznthou — tznthou.com · [email protected]