claude-mem

manymems

Team memory for multi-session, multi-modal agentic coding agents.

This project extends claude-mem — a persistent memory plugin for AI coding tools — into a shared memory system for engineering teams. Two capabilities drive the work:

• Intent → Code Provenance. Link every developer prompt to the exact lines it changed, anchored by tree-sitter symbols so the link survives later edits. Ask "why was this written?" and get the original intent back.
• Team multi-session memory. Let Claude Code, Cursor, Codex, Windsurf, and other agents share a common memory pool with identity, privacy controls, conflict detection, multi-modal capture, and team-scoped retrieval.

---

Why this project exists

AI coding agents remember nothing across sessions and share nothing across teammates. Each developer rediscovers the same patterns, re-explains the same context, and loses the reasoning behind every change the moment the session ends.

claude-mem solves the single-developer case. manymems solves the team case.

---

What claude-mem already provides

Developer prompt
      │
      ▼
UserPromptSubmit hook
      │  stores prompt_text in user_prompts table
      ▼
PostToolUse hook  ──►  observation handler
      │                      │
      │                      ▼
      │              Claude Agent SDK
      │              (compresses tool events
      │               into structured observations)
      │                      │
      ▼                      ▼
  SQLite ◄──────────  storeObservation()
  (local)              files_modified: [paths]   ← paths only, no line ranges
      │
      ▼
Chroma (vectors)
      │
      ▼
SessionStart hook
  injects relevant past observations
  into the next session's context

The team server (Postgres + BullMQ + better-auth, merged in PR #2383) adds multi-user identity, API-key auth, and a horizontally scalable worker. The substrate is sound; the gaps are provenance and team-layer features.

---

What manymems adds

Track A — Intent → Code Provenance

User types a prompt
      │
      ▼  (UserPromptSubmit)
user_prompts row  ──── prompt_id ────────────────────────┐
                                                          │
      │                                                   │
      ▼  (PostToolUse: Edit / Write)                      │
A1: extract line ranges                                   │
      │  new_string found in file after edit              │
      │  line_start, line_end computed deterministically  │
      ▼                                                   │
A2: build symbol anchor  (tree-sitter)                    │
      │  qualified_name: "AuthService.login"              │
      │  signature_hash: sha256(signature)  ← staleness   │
      ▼                                                   │
A3: code_provenance row ◄─────────────── prompt_id ───────┘
      │  file_path, line_start, line_end
      │  symbol_qualified_name, signature_hash
      │  old_content_hash, new_content_hash
      │  commit_sha  (backfilled by A4)
      ▼
A4: git post-commit hook
      │  fills commit_sha on matching rows
      ▼
A5: get_code_provenance MCP tool
      │  input:  file, line
      │  output: [{prompt_text, symbol, line_range,
      │            commit_sha, stale}]
      ▼
  "why was this written?"  answered

Staleness: when a query arrives, manymems re-hashes the symbol's current signature. If it differs from the stored signature_hash, the record returns stale: true.

---

Track B — Team Multi-Session Memory