Co-Engram/Co-Engram
Self-evolving team memory as plain Markdown in git: a native Cordis plugin that registers 38 bare-name memory tools plus a prompt-signals section re-evaluated at every assembly, with RPE reinforcement, decay and sleep consolidation; shares one data repo with its Claude Code (MCP) and OpenClaw hosts; verified against DSH 0.1.0-rc.6.
A self-evolving memory system for AI agents and teams that models memory after the brain rather than a vector store: engrams strengthen with use, weaken when they fail, consolidate during sleep, and verify themselves through metacognition. Every memory is one Markdown file with YAML frontmatter under a ULID that never changes, so renames and moves do not break references while Git diffs stay clean; connections between memories are independent per-edge synapse files keyed by a deterministic hash of (from, to, kind). A maintenance engine runs light (RPE-based reinforcement), deep (consolidation + decay) and rem (metacognition upgrade/refute) stages automatically, and implicit memory proposals pass a two-layer filter (a zero-cost rule-based prefilter plus a necessity evaluator) so mechanical repetition is rejected and only genuinely reusable decisions become candidates. The README frames the whole thing as working through conversation — the agent decides when to capture, search or update.
Install
dsh plugin --profile <name> add @co-engram/dshREADME DeepSeek Harness section quoted verbatim (replace <name> with your profile). Registry-verified live 2026-09-13: @co-engram/dsh 0.6.0 with a matching repository field (github.com/Co-Engram/Co-Engram). The README states that is all that is needed — the package ships a dsh.bundle patch, so the plugin activates as a profile layer with zero manual config, all 40 tools register under bare names (engram_search, …), and a memory:co-engram prompt section injects live signals (top tags / skills / path overview) at every prompt assembly. The same data repo is shared with the Claude Code and OpenClaw hosts through a process lock. The Claude Code host installs separately with npm install -g @co-engram/claude-code; the OpenClaw host with openclaw plugins install @co-engram/openclaw --dangerously-force-unsafe-install.
Compatibility
A host-agnostic TypeScript core with host integrations for Claude Code (via MCP), OpenClaw (via the plugin SDK) and DeepSeek Harness (via a native Cordis plugin). The quickstart initializes the memory data repo as a separate Git repository (~/team-memory), pointed at with co-engram config data-root, and hosts share it through a process lock.
Details
- Repo: Co-Engram/Co-Engram
- Category: Memory
- Stars: 10
- Version: npm @co-engram/dsh 0.6.0 (registry-verified 2026-09-13); latest GitHub release v0.1.5
- Last push: 2026-08-22
- First seen: 2026-06-26
Recent updates
The latest GitHub release is v0.1.5, titled "Trust boundary redesign + fail-loud + doctor dangling-ref cleanup"; the README does not publish a release-by-release changelog table, so verify current versions on the registry and the release page.
FAQ
- How do I install Co-Engram for DeepSeek Harness?
- The README's DSH section is a single command: dsh plugin --profile <name> add @co-engram/dsh. The bundle patch activates it as a profile layer, registering all 40 tools under bare names.
- Does it work with other hosts?
- Yes — the same core has Claude Code (MCP) and OpenClaw (plugin SDK) hosts, and the README says the hosts share the same data repo via a process lock.
- What makes it different from a vector store?
- Per the README it models engrams: they strengthen with use, weaken when wrong, consolidate during sleep, and verify themselves through metacognition — plus stable ULID file IDs so renames and moves never break references.
Alternatives
csyangwen/dsh-memory-evolve · Towzai/dsh-memory · Scorp1o117/dsh-tdai-memory