Your AI coding agent needs a hippocampus

What it does

Engram is a single Go binary that stores memories for AI coding agents in a local SQLite database with FTS5 full-text search. It exposes nineteen MCP tools—save, search, session lifecycle, and conflict surfacing—so any MCP-compatible agent can persist context across sessions. Humans can browse the same memory through an interactive TUI or HTTP API.

The interesting bit

The project treats agent memory like a real engram—a physical trace—by indexing observations with structured fields and surfacing conflicts when an agent contradicts itself. A beta feature even pipes FTS5 candidates through your existing Claude Code or OpenCode CLI to judge semantic similarity, turning lexical search into an LLM-mediated sanity check.

Key highlights

• Agent-agnostic via MCP stdio: works with Claude Code, Cursor, Windsurf, Gemini CLI, Codex, and others.
• Runs fully offline with zero dependencies; local SQLite is always the source of truth, while cloud replication and Git sync are strictly opt-in.
• Nineteen MCP tools including conflict surfacing (mem_judge, mem_compare) and a repair/doctor flow for cloud sync issues.
• Interactive TUI with vim-style navigation and Catppuccin Mocha theme for manual memory inspection.

Caveats

• Cloud sync can hit canonicalization failures that require manual doctor/repair cycles; the tool deliberately never auto-applies fixes.
• Updating the binary while an MCP client is running leaves the old stdio subprocess in place until you restart the client.
• The semantic conflict-judge feature is beta and leans on your existing Claude Code or OpenCode CLI subscription to run LLM reasoning.

Verdict

Worth a look if you juggle multiple AI agents and want a single, local memory store that survives reboots. Skip it if you already have a cloud-native memory layer you trust or rarely work across multiple sessions.