The claim every memory tool makes is that your data stays yours; this one ships the check. Memory Lens renders a read-only report of precisely what a caller would receive and what was withheld, with secrets shown redacted, so "local and inspectable" is something you verify rather than accept. The files under `~/.engram/` are plain JSON and Markdown for the same reason.
A local-first store of your working identity — preferences, quality bar, lessons, decisions and project context — that any MCP-compatible tool can read. The distinction it draws is deliberate: it is not an agent memory database holding task history, it is the user-owned layer above whatever memory Claude Code, Codex, Cursor or Windsurf each keep for themselves.
- A session-start call that loads identity plus relevant knowledge, with a `token_budget` parameter so you control how much context it spends — `get_user_context`, `get_recall`
- Three knowledge types — lessons, decisions and playbooks — written through one unified endpoint that routes by kind — `memory_store`, `add_lesson`, `add_decision`, `add_playbook`
- Search across all three with domain, tier and date filters, plus a project-relevant view — `search_knowledge`, `get_relevant_knowledge`
- Session recovery after a restart and a cross-tool resume brief — `get_recent_context`, `get_resume_brief`, `get_daily_log`, `wrap_up_session`
- Project snapshots saved and read back so a new session starts where the last one stopped — `save_project_snapshot`, `get_project_context`
- A Markdown identity card exported for tools that do not speak MCP at all, such as ChatGPT or Gemini — `get_identity_card`
- Memory Lens: `engram preview --html` shows exactly what an AI caller would receive and what governance withheld — including a lesson containing a credential shown as `[REDACTED]` — before anything is sent
- An advanced surface behind `ENGRAM_TOOLS=all` covering knowledge review and staging, merging and relations, graph exploration, import and export, caller trust management and the audit log — `review_staging`, `merge_knowledge`, `explore_knowledge`, `manage_caller_trust`, `get_audit_log`
- A self-diagnosis tool for when the store misbehaves — `doctor`
Python 3.10 or higher: `pip install piia-engram && engram setup`. The wizard detects Claude Code, Cursor, Codex and Claude Desktop, lists the exact config files it will touch, backs each one up, and writes only after a confirm — decline and nothing changes. Identity and knowledge live under `~/.engram/` as plain JSON and Markdown. High-risk items — credentials, shell commands, MCP config, permission rules — wait for your review; low and medium writes are absorbed automatically but stay auditable and reversible, and `ENGRAM_APPROVAL=strict` gates every write. `ENGRAM_TOOLS=all` unlocks the advanced surface; `ENGRAM_MCP_STARTUP_SYNC` controls whether startup reconciliation runs in the background, eagerly, or not at all. Identity and knowledge tools make no network calls by default — the exception is `read_web_content`, which fetches a URL you supply. Licensed AGPL v3.
One command — pip install piia-engram && engram setup