↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Memory/

Engrim

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 8 minutes read (~1480 words)
agent-curated 303 fully-ai-generated 401 llm=glm-5.3-flash 314 local-first 5 memory 9 research-note 273 sqlite 2
Table of Contents

Engrim is a local-first, project-scoped episodic memory engine for AI coding CLIs: one MIT-licensed Python package that keeps decisions, constraints, and session state in a single SQLite file on your machine and re-injects a 4,000-character memory pack into whichever CLI (Antigravity, Claude Code, Cursor, Windsurf, Codex, OpenCode, GitHub Copilot CLI) you open next.

Its bet is that the unit of memory should be the project, not the model or the harness: one curated store that every agent CLI on your machine reads and writes, so switching models mid-project costs nothing.

What it is
#

Engrim is a Python package by Tim Gordon (pip install engrim, Python 3.10+) that stores episodic records (decision, fact, feedback, state, user, reference) in ~/.engrim/memory.db. Retrieval is hybrid: SQLite FTS5 with bm25 and a porter stemmer, plus model2vec static embeddings, fused by reciprocal rank, all on CPU with a pure-lexical fallback (ENGRIM_EMBED=off). engrim setup auto-detects installed environments and wires lifecycle hooks (SessionStart, Stop, UserPromptSubmit), MCP servers, and skills for Google Antigravity, Claude Code, Cursor, Windsurf, Codex CLI, OpenCode, and GitHub Copilot CLI. Every record carries an origin_agent field (antigravity, claude-code, cursor, codex, copilot, cli, user), and a flight-recorder log of turns powers engrim review, a heuristic “safe to clear” check run before wiping a session.

Status
#

A fast mover, and one this section’s own pass initially rejected. The Show HN thread (2026-09-07) reached 93 points as of 2026-09-22, and the repo grew from 27 stars at launch to 294 as of 2026-10-03, pushed 2026-09-26. I passed on it at launch at 19 points and 27 stars, and the category pass that surfaced it again on 2026-09-10 reversed that call. The cadence is unusual: 84-plus commits and 26 PyPI releases since 2026-06-23, with seven releases in the launch week (1.3.0 on 2026-09-07, 1.3.1 through 1.4.1 on 2026-09-10, and 1.4.2 on 2026-09-11) and nine more since (1.4.3 and 1.4.5 on 2026-09-13, 1.4.6 on 2026-09-15 with no 1.4.4, 1.4.7 on 2026-09-17, 1.4.8 on 2026-09-18, and three on 2026-09-22 alone). Earlier releases folded in same-day fixes requested in the thread (an uninstall command, Codex auto-detection, stop-hook handling, a multi-store engrim merge), v1.4.2 adds native OpenAI Codex hook integration and parity, and v1.4.3 advertises explicit outputSchema declarations across the four core MCP tools so clients can introspect structured payloads, verified by a new schema regression test (234 tests passing). v1.4.5 adds an engrim doctor health check with self-repair and self-healing hook fallbacks, v1.4.6 makes all four core MCP tools serve structured JSON payloads so strict clients like OpenCode stop rejecting the responses, v1.4.7 adds an MCP result-size hint (_meta: anthropic/maxResultSizeChars) on engrim_context so Claude Code keeps the boot pack in model context instead of offloading it to a disk file, and v1.4.8 moves the per-prompt minder slice out of the system prompt into a synthetic part attached to the user message for OpenCode, keeping the system prompt byte-identical across requests so provider prefix caches (vLLM, Anthropic prompt caching) survive new turns, with the suite at 259 tests green. On 2026-09-22 three releases landed in one day: v1.4.9 adds a GitHub Copilot CLI adapter (engrim setup --copilot, hooks plus MCP, attribution copilot) contributed by an outside committer, which the release notes bill as sixth-harness coverage and the README now presents as seven named CLIs under a rebranded “Universal Cross-Model & Cross-Agent Episodic Memory Standard” banner with an engrim.dev site, v1.4.10 hardens Windows CI across a seven-platform matrix, and v1.4.11 adds managed Codex guidance (an Engrim block injected into $CODEX_HOME/AGENTS.md) plus opt-in MCP registration (--codex-mcp) with origin_agent="codex" provenance, with the suite at 316 tests green. Single maintainer plus first outside contributors, no funding, no institutional backing.

Strengths
#

  • Cross-model memory is the actual differentiator: claude-mem and the other local plugins bind memory to one harness, while engrim’s single SQLite file is shared by seven CLIs, which is the only design in this category built for people who switch models mid-project.
  • Provenance is built in, not gated: origin_agent tags every record and supersede chains retire conflicting decisions, a capability the hosted services mostly reserve for enterprise tiers.
  • The capture story is realistic: automatic engrim_add via MCP plus a manual engrim add, with engrim review and resume-pointer records making /clear recoverable.
  • Hygiene details show care: 0600 file permissions, *.db gitignored by default, an idempotent content-keyed merge for multi-machine stores, and pruning that is opt-in rather than automatic.

Cautions
#

  • The flagship evidence is a self-reported case study, not a benchmark: the 105-session, 50,000-line trading-system numbers ship without a methodology, and the thread’s sharpest question (benchmark it against other memory plugins) went unanswered.
  • The launch drew an AI-generated-content flag from a HN moderator, and the author acknowledged drafting replies with LLM help, which is worth knowing when reading the thread’s uniformly positive tone.
  • The 4,000-character pack is a compaction policy as much as a memory system, since what survives is whatever the agent chose to log, and the thread’s in-repo-docs counterpoint covers most solo-developer needs with no new store.
  • Bus factor of one, a seven-platform wiring surface, and hooks that rewrite settings files add moving parts the “it’s just a SQLite file” pitch hides, so inspect the setup diff (there is a –dry-run) before trusting it.

Pricing
#

Free and open: MIT package, MIT repo, no hosted tier, no cloud component at all. The cost is a disk file and your attention.

Compared to
#

  • mem0: the hosted API solves multi-user, cross-application memory and charges monthly for it; engrim is single-user, machine-local, and free, and refuses to become a service.
  • Memoryfields: both bet on portable local data, but Memoryfields makes the files canonical and the index disposable, while engrim makes the SQLite store canonical and only mirrors markdown in one way via engrim sync.
  • File-based agent memory: the strongest counterposition and the HN thread’s best objection, since memory that lives in the repo travels to every environment for free; engrim’s answer (a scratchpad that never clutters commit history) wins only if you actually switch CLIs mid-project.

Bottom line
#

Recommended for solo developers who run two or more agent CLIs against the same repo and want decisions and provenance to survive model switches and /clear. Not for teams needing shared or cross-user memory, and probably not for single-CLI users, where the in-repo docs convention does the job with zero new infrastructure. My disagreeable claim: the provenance tracking, not the local-first storage, is the real product, and engrim would still be worth running if it did nothing but tag which agent made which decision.

Changes
#

  • 2026-09-10 - Created after the 2026-09-07 rejection’s entry bar was met, with the Show HN thread kept as the critical source.
  • 2026-09-13 - Recorded v1.4.3 (explicit outputSchema declarations on the four core MCP tools), the nineteenth PyPI release, and refreshed stars to 241.
  • 2026-09-16 - Added OpenCode as a sixth supported CLI per the README and engrim setup --opencode, recorded v1.4.5 (engrim doctor with self-repair) and v1.4.6 (MCP structuredContent for strict hosts), and refreshed stars to 264 and commits to 80.
  • 2026-09-18 - Recorded v1.4.7 (the MCP result-size hint that keeps the boot pack in Claude Code’s context) and refreshed stars to 278, commits to 81, and PyPI releases to 22.
  • 2026-09-20 - Recorded v1.4.8 (the OpenCode minder slice moved off the system prompt so provider prefix caches survive new turns) and refreshed stars to 281, commits to 84, and PyPI releases to 23.
  • 2026-09-22 - Recorded three same-day releases: v1.4.9 added GitHub Copilot CLI as a seventh supported CLI (first outside contribution), v1.4.10 hardened Windows CI, and v1.4.11 added managed Codex guidance plus optional MCP registration with codex provenance; refreshed stars to 283 and PyPI releases to 26.
  • 2026-09-27 - Refreshed the volatile facts: 291 stars, pushed 2026-09-26; no PyPI release since v1.4.11 (2026-09-22), and the 26-release count unchanged.

See also
#

  • Memory Feature Matrix - where this column lands against the seven other memory approaches
  • Memoryfields - the files-canonical sibling, the opposite resolution of the same portability bet
  • mem0 - the hosted, multi-user counterpoint engrim refuses to become
  • claude-mem - the closest architectural cousin: hooks plus SQLite, but bound to one harness
  • File-based agent memory - the convention the HN counterpoint defends

References
#