Piia Engram is a local-first AI work identity and memory layer from an individual-led open-source project (AGPL-3.0), not an Anthropic or MCP-official offering. Its positioning is deliberate and clearly differentiated from agent task-memory stores such as Mem0, Zep, and Letta: it stores who you are — identity, preferences, quality standards, lessons, and decisions — rather than what happened in one workflow. Data lives as plaintext JSON/Markdown under ~/.engram/, with zero network calls by default and no cloud account, so you can open, edit, back up, or migrate it yourself. Nineteen core tools load by default, with ENGRAM_TOOLS=all unlocking all 59; writes pass through a tiered governance and staging model where high-risk content needs human approval and strict mode can gate everything. Cross-tool continuity is backed by a verifiable evidence path (a documented Claude Code → Codex continuity proof plus a synthetic MCIC benchmark), and a read-only Preview/Memory Lens shows exactly what a caller would receive. Its boundaries are stated just as honestly: plaintext storage by default, caller identity from environment variables rather than strong authentication, governance as a local policy layer rather than a hardened sandbox, and an explicit warning not to store keys or customer PII. A strong fit for individual developers who juggle multiple MCP coding tools and value local data sovereignty; a poor fit for teams needing shared memory, enterprise support, or a commercial license.
Piia Engram is a local-first AI work identity and memory layer for MCP-compatible coding tools such as Claude Code, Codex, Cursor, Windsurf, and Claude Desktop. You tell it once who you are, how you work, and what "good" means; multiple AI tools can then start from the same approved context — identity, preferences, quality standards, code-review bar, lessons learned, key decisions, and project snapshots — stored as plain JSON and Markdown under ~/.engram/ on your machine. There is no cloud account and no hidden memory you cannot inspect.
It deliberately occupies a different layer from agent task-memory databases like Mem0, Zep, or Letta: those store what happened during a workflow, while Piia Engram stores who is behind every task. It ships 59 MCP tools, but only 19 Tier-1 core tools are loaded by default to keep context clean; setting ENGRAM_TOOLS=all unlocks the full surface (knowledge management, governance, import/export, admin).
Writes follow a governance model: AI writes locally, high-risk items (credentials, shell commands, MCP config, permission rules) wait in staging for your review, and low/medium writes are auto-absorbed but fully auditable and reversible. ENGRAM_APPROVAL=strict gates every write. Other capabilities include cross-project knowledge inheritance, session-insight extraction, playbook auto-extraction (drafts only — you confirm before a playbook becomes trusted), a local tools registry, knowledge health and near-duplicate discovery, optional hybrid retrieval (FTS5 plus a semantic vector layer, off by default), optional field-level AES-256-GCM encryption, and a local audit log.
By default the identity and knowledge tools make zero network calls (except the optional read_web_content); telemetry is off by default, and remote telemetry or feedback requires separate explicit opt-in and sends counts only. Licensed AGPL-3.0, created by @Patdolitse with Claude Code and Codex as AI assistants — a community open-source project, not an Anthropic or MCP-official offering.
Option 1 (pip): run pip install piia-engram && engram setup. The wizard detects your Python environment, lets you choose the Engram data folder (~/.engram, another drive, or a custom path), detects your AI tools and lists the exact config files it will touch, and writes the MCP connection only after a one-keystroke confirm (every write is backed up first; decline and nothing changes). It then walks you through seed knowledge (role, tech stack, language), can smart-import rules from existing CLAUDE.md / .cursorrules files, and previews your AI identity card.
Option 2 (uvx, no prior install): use uvx --from piia-engram piia-engram-mcp in your client config.
Client config example (Claude Desktop / generic MCP over stdio):
{
"mcpServers": {
"piia-engram": {
"command": "python",
"args": ["-m", "piia_engram.mcp_server"]
}
}
}Cursor uses ~/.cursor/mcp.json, Codex uses ~/.codex/mcp.json, Claude Desktop uses claude_desktop_config.json. Claude Code also supports claude mcp add piia-engram -- piia-engram-mcp.
For non-interactive or CI runs, engram setup --apply-external-config skips the confirmation prompt (backups are still taken). Restart your AI tool afterwards and run engram doctor to verify connectivity.
Optional environment variables: ENGRAM_TOOLS=core|all (default core exposes the 19 core tools; all unlocks the full 59), ENGRAM_MCP_STARTUP_SYNC=off|background|eager (default off; background is for local cross-tool sync), PYTHONIOENCODING=utf-8 (Windows-style consoles).
Remote deployment: pip install piia-engram[remote], generate a token, then start ENGRAM_AUTH_TOKEN=... python -m piia_engram.mcp_server --transport sse --host 0.0.0.0 --port 8767. Clients connect with a url plus an Authorization: Bearer header; always put it behind TLS in production.
{
"mcpServers": {
"piia-engram": {
"command": "python",
"args": ["-m", "piia_engram.mcp_server"]
}
}
}