Changelog

All notable changes to mememory are documented here.

v0.6.0 — 2026-04-27

Pinned delivery + forced recall

• New pinned delivery type. Reinjects on every agent turn through the UserPromptSubmit hook, wrapped in a <system-reminder> block with rotated framing. Where bootstrap loads once at session start and drifts up the context, pinned reasserts itself every turn so hard rules don't decay into background. See Pinned Rules & Forced Recall.
• System meta-rules layer is hard-coded in mememory and prepended to every pinned payload — directives about working with the memory layer itself (recall mandate, code-vs-memory truth source, "rule violation = task failure"). Five formulations of each meta-rule rotate per render, plus rotated opening/closing imperatives, to defend against agent adaptation to a single phrasing.
• Forced recall via PreToolUse hook. A new lock file (${TMPDIR}/mememory-recall-pending-${session_id}) is armed at SessionStart and removed by PostToolUse on mcp__mememory__recall. While the lock exists, mememory recall-gate denies any tool call that isn't an mcp__mememory__* tool — the agent physically cannot work on the user's task without first loading project context via recall. Stale locks (older than 24h) are garbage-collected at the next SessionStart.
• One-shot installer: mememory install-hooks patches ~/.claude/settings.json with all four hooks (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse). Idempotent, preserves existing settings and foreign hooks, and writes a timestamped backup before any change. --uninstall removes them cleanly.
• CLI: new commands pinned, recall-gate, recall-ack, install-hooks. mememory setup now interactively offers to install the hooks at the end of its run.
• MCP resources: mememory://pinned and mememory://pinned/{project}.
• Soft budget warning at 5,000 tokens for the pinned payload — informational, never blocks. Pinned must stay tight to act as a checklist; the warning surfaces when the payload is starting to dilute that effect.

Admin UI

• Pinned Preview page at /pinned renders the exact payload your agent receives for any project, with token estimates and counts of pinned memories at each scope.
• Delivery filter on the Memories page now includes pinned. The New Memory form lets you create pinned entries directly.
• Persona scope cleanup. The legacy persona scope (removed from the backend long ago) is no longer surfaced anywhere in the admin UI — filter, form, badges, settings, search filter, and the leftover CSS token are all gone.

Notes

• Codex CLI install-hooks parity is scheduled for the next release. Codex users can still wire SessionStart → mememory bootstrap --hook manually using the existing instructions; full pinned + forced-recall integration follows in 0.7.0.
• Backward compatibility. No SQL migration. The delivery column is TEXT, so adding a new value is a Go-only change. Existing bootstrap and on_demand memories are untouched.

Unreleased

Refactor: pluggable connection

• Renamed internal/memory package to internal/engine to better separate domain logic from the "memory" concept.
• Removed hardcoded defaultDatabaseURL constants from server binaries. DATABASE_URL is now required; missing value causes fail-fast with a clear hint.
• Added pgvector preflight check in pg.NewClient — server fails fast with installation hint if the extension is missing.
• Renamed MEMEMORY_DATA_DIR env var to DATA_DIR.
• Added OS-aware default DATA_DIR resolver in the mememory CLI: ~/.local/share/mememory (Linux/XDG), ~/Library/Application Support/mememory (macOS), %LOCALAPPDATA%\mememory (Windows).
• New mememory setup command: resolves DATA_DIR, writes .env, brings up bundled Docker stack.
• New mememory uninstall command: stops containers but preserves data by default. --purge flag requires interactive path confirmation to delete data.
• PostgreSQL password default changed from memory to mememory.

Added

• bootstrap memory type: only memories of this type are loaded automatically at session start. All other types are fetched on demand via recall.
• Bootstrap System section: hard-coded directives at the top of every bootstrap output — "use mememory as the only persistent memory source" and "always call recall on the user's first message".
• Bootstrap size limit: output is capped at 10KB (MaxBootstrapBytes). mememory bootstrap prints a stderr warning when the limit is exceeded, and remember(type="bootstrap", ...) warns when the combined set would exceed it.

Changed

• Two scopes only: removed the persona scope. Scopes are now global and project. The --persona CLI flag, persona column in filters, by_persona stats field, and ScopePersona constant are gone.
• Updated scope weights in recall scoring: project=1.0, global=0.8 (previously persona=1.0, project=0.8, global=0.6).
• Bootstrap loads only type=bootstrap: mememory bootstrap, memory-server --bootstrap, and the memory://bootstrap / memory://bootstrap/{project} MCP resources now filter by type=bootstrap instead of returning all memories in scope.
• MCP config path: Claude Code configuration moved from ~/.claude/.claude.json to ~/.claude/.mcp.json (or project-level .mcp.json). The server name is now mememory (was memory).
• MCP server instructions advertise two scopes (global, project) and six types (fact, rule, decision, feedback, context, bootstrap).

Renamed

• Renamed binaries: memory-server → mememory-server, memory-admin → mememory-admin.
• MCP resource URI scheme renamed from memory:// to mememory://. The bootstrap resources are now mememory://bootstrap and mememory://bootstrap/{project}.
• Environment variable MEMORY_DATA_DIR renamed (later superseded by DATA_DIR — see "Refactor: pluggable connection" above).

Breaking

• PostgreSQL database and user renamed from memory to mememory. Existing installations must reset the postgres data directory (stop the stack, move $DATA_DIR/postgres aside, then mememory setup) — the old memory database will not be migrated automatically.

Features

• MCP server with 7 tools: remember, recall, forget, update, list, stats, help
• MCP resources: memory://bootstrap and memory://bootstrap/{project} for session initialization
• Hierarchical scopes: global and project, with project automatically inheriting global during recall
• 6 memory types: fact, rule, decision, feedback, context, bootstrap
• Semantic search via PostgreSQL pgvector with cosine similarity and HNSW indexing
• Composite scoring: similarity x scope_weight x memory_weight x temporal_decay
• Contradiction detection: warns when new memories are >75% similar to existing ones
• Belief evolution: supersedes parameter for clean knowledge chains with auto-downgrade
• TTL support: auto-expiring memories with hourly cleanup
• Pluggable embedding providers: Ollama (local, default) and OpenAI (+ any OpenAI-compatible API)
• Dynamic embedding dimensions: auto-detected at startup, validated against database
• Admin API: full REST API for CRUD, search, bulk delete, export/import
• Admin web UI: React SPA for browsing and managing memories
• mememory CLI: native Go binary with bootstrap, status, version commands
• Session bootstrap: Markdown-formatted output of bootstrap-type memories for SessionStart hooks, capped at 10KB
• Auto-detect project: CLI detects project name from git repository root
• Docker stack: PostgreSQL (pgvector) + Ollama + Admin in a single docker compose up
• Privacy first: all data stays local, no external telemetry