Engram Memory Skill
Give your AI coding agent persistent cross-session memory that proactively saves decisions, conventions, bugs, and discoveries.
Evidence shows the skill instructs proactive saving of user decisions, preferences, and other sensitive information without explicit user consent, posing over-collection risks; however, it provides local storage, opt-in cloud sync, and a security policy, making risks visible. Deductions: lacks user confirmation or permission management mechanisms, data sharing and retention policies unclear; needs least privilege and transparency improvements.
Static review cannot verify tool availability or actual functionality; relies on external MCP server and hooks, with no detailed error handling or fallback. Deductions: no reproducible test evidence, documented steps unverified, error handling insufficient; score capped at 10.
Skill clearly targets agents like Claude Code with clear scenarios, but actual trigger accuracy and boundary conditions unverified, non-fit scenarios not specified. Deductions: trigger conditions documented but no empirical support; environment fit and Chinese-language support not mentioned.
Documentation is well-layered with license and installation notes, but lacks version history, changelog, and clear maintenance responsibility. Deductions: maintenance and update path unclear; good information architecture but hidden assumptions exist.
Skill claims to substantially improve memory persistence, but static evaluation cannot confirm actual results, and relies on user following instructions; value proposition plausible but comparative benefit over manual memory lacks evidence. Deductions: output format and completeness unverified; value unproven; score capped at 7.
Only source files and CI config present, but no independent verification or reproducible test results. Deductions: evidence limited to single source, no cross-validation, claims cannot be independently verified; score capped at 5.
- The skill proactively records user decisions and preferences, which may include sensitive information; users should review content before sharing or backing up memory.
- Depends on external MCP hooks and server; network or environment changes may cause tools unavailable; ensure correct installation and configuration.
- Static evaluation does not verify actual runtime; test in isolated environment before production use.
What does this skill do, and when should you use it?
Engram is a persistent memory system for AI coding agents, exposed via MCP. This skill is a Claude Code SKILL.md that enforces a mandatory memory protocol: the agent must proactively call tools like mem_save after every decision, bug fix, discovery, or user confirmation. It defines search procedures, session summary requirements, and recovery after context compaction. By using this skill, agents retain context between sessions, reducing redundant work and preventing knowledge loss.
This skill enforces a mandatory memory protocol that includes: proactively saving decisions, conventions, bugs, discoveries, and user preferences; using mem_save and other tools to persist memories into Engram's SQLite database with FTS5 full-text search; searching memory via mem_context then mem_search when recall is needed; requiring mem_session_summary at session end or before compaction; and instructing to re-run engram setup claude-code if tools are missing. The skill itself doesn't execute processes but relies on the Engram binary's MCP tools.
- A developer works on a project across multiple sessions, needing the agent to remember architectural decisions and coding conventions
- After fixing a complex bug, the developer wants the agent to save the root cause to avoid re-diagnosis
- During a session, the agent learns user preferences (e.g., coding style) and automatically saves them
- After a context compaction, the agent recovers context via a session summary to continue unfinished work
What are this skill's strengths and limitations?
- Proactive memory: agent saves critical info without user prompting
- Cross-session persistence: survives context compaction via summaries and search
- Complete toolset: provides full MCP tools (save, search, update, etc.)
- Compatibility: designed to work with various MCP clients (Claude Code, OpenCode, etc.)
- Requires external Go binary installation and configuration
- No documented test suite; reliability not fully proven
- Mandatory protocol may add overhead to agent operations
- Some documentation lacks specifics (e.g., exact search commands)
How do you install this skill?
- Install Engram:
brew install gentleman-programming/tap/engram(or other methods) - Setup Claude Code:
claude plugin marketplace add Gentleman-Programming/engram && claude plugin install engram - Restart Claude Code to load MCP tools.
How do you use this skill?
The skill is always active; no manual trigger needed. The agent proactively saves memories. When recall is needed, just ask; the agent will invoke mem_context first, then mem_search. The skill also requires the agent to summarize memories at the end of a session.
How does this skill compare with similar options?
The skill does not explicitly compare to other memory systems, but mentions inspiration from claude-mem.