What does this skill do, and when should you use it?
This protocol is mandatory and always active, requiring the agent to proactively save architecture decisions, bug fixes, user preferences, and discoveries into Engram's SQLite store. It leverages an MCP server for tools like mem_save and a session-close summary protocol to ensure continuity. Installation involves running `engram setup codex` and restarting Codex; tools auto-load via the UserPromptSubmit hook.
- Calls
mem_saveimmediately after relevant events (decisions, completed work, discoveries, user confirmations/rejections) with structured fields (title, type, What/Why/Where/Learned). - Provides
mem_contextandmem_searchfor retrieval of past memories. mem_session_summarylogs a summary at session end.- After compaction, it persists a summary via
mem_session_summaryand recovers context withmem_context. - All actions go through the MCP server.
- A developer fixes a stubborn bug and wants the root cause remembered next session.
- A team establishes a coding convention and expects the agent to enforce it consistently.
- A user expresses a preference for a tool or framework, and future work should follow it.
- After context compaction, quickly restore the session's state.
- At the start of a new session, the user asks about prior work; the agent auto-searches memory.
- Environments without the engram binary or where MCP setup fails — tools will be unavailable
- Scenarios requiring memory access over HTTP or Docker networks — MCP-only, not reachable remotely
- Mission-critical production use needing stability guarantees — no test suite and relies on agent adherence to the protocol
How do you install this skill?
- The skill is designed to be always active and mandates memory saving and session-end protocols; administrators should ensure users are aware of this behavior and configure permissions as needed.
- Default HTTP server is unauthed; consider setting ENGRAM_HTTP_TOKEN to protect local data if exposed.
- The skill does not explicitly prohibit saving sensitive information; ensure user data is not stored in memory unless explicitly consented.
- The skill relies heavily on the external engram binary and MCP server; ensure these components remain updated and available to avoid disruption.
- The beta features should be used with isolated test environments to avoid impacting production.
- Shell / CLI
- Local filesystem
- MCP Server
engram binary (Go, no runtime deps)
Codex
brew install gentleman-programming/tap/engram
engram setup codexRestart Codex after setup. Ensure MCP tools are available; if missing, rerun engram setup codex and restart. Other platforms per docs.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- What did we decide earlier? Recall what we did last session
- How did we solve this bug last time?
- Save this architecture decision to Engram
- Continue where we left off last session
No manual activation is needed—the protocol auto-runs. The agent triggers mem_save on relevant events and mem_session_summary at session end. Users can prompt with phrases like 'remember how we solved this last time?' to trigger search.
What are this skill's strengths and limitations?
- Proactive saving reduces forgetting
- Preserves context across sessions and compactions
- Structured memory with topic keys and upsert
- Automatic tool loading—no manual search
- Supports project and personal scopes
- Depends on external engram binary and MCP config; setup failures break tools
- Relies on agent adherence; otherwise memory may be incomplete
- MCP-only, not reachable over HTTP or Docker networks
- No test suite or evidence of stability
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Engram Persistent Memory Protocol this page | 68 · Recommended | ★ 7.1k | 4d ago | MIT |
| Engram Memory Skill | 54 · Use with care | ★ 7.1k | 4d ago | MIT |
| Engram Architecture Guardrails | 50 · Use with care | ★ 7.1k | 4d ago | MIT |
| Engram Business-Rules Guardrails | 54 · Use with care | ★ 7.1k | 4d ago | MIT |
| agentMemory — Hybrid Memory System for AI Coding Agents | 42 · Not recommended | ★ 33 | 8mo ago | MIT |
How did FollowSkills review this skill?
The skill requires the agent to proactively save memories and enforces a session-end protocol without requiring explicit user confirmation, and lacks explicit guidance on handling sensitive data in memory storage. The skill exhibits least privilege in terms of permissions; the MCP tools and HTTP server data flow are visible, but the HTTP server defaults to being open if token is not set, which could expose data. Installation and setup info provided, attribution clear. Deductions: lacks user confirmation or sensitive-data transparency.
The skill provides a clear protocol with mandatory actions and fallback (rerun setup), but no explicit error handling or failure-feedback details for abnormal inputs. README mentions tests and CI workflow, offering some evidence, but static review prevents full verification. Deductions: lack of error messages or test coverage evidence.
Targets AI coding agents with persistent memory needs; scenario and boundaries are clear, triggers defined, non-fit cases evident. It works with Chinese and English languages and runs locally, so no internet dependency. Deductions: limited evidence beyond default scenarios.
Documentation is well-layered: README, DOCS, INSTALLATION, etc. License is MIT. But no version or changelog information provided. Deductions: missing changelog or versioning.
Goals and benefits are clear; the skill provides detailed instructions and tool usage. It likely completes the core task of persisting memory, but direct usability or output verification is not proven without execution. Deductions: no execution evidence.
CI workflow with go test ./... and e2e tests provides some verifiability, but static review limits score. No independent reproduction performed. Deductions: no independent execution evidence.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →