Engram — brain-inspired long-term memory — for opencode. Injects relevant memory at session start, consolidates when a session goes idle, adds /engram-* commands, and recalls on demand via a skill. One Rust binary, zero deps, memory store shared with the Claude Code & Codex versions.

English | 中文

opencode plugin engram-opencode

That's it — opencode installs the npm package and adds it to your config. On first run the plugin self-provisions everything via its config hook: the pre-permissioned reviewer/query agents and the /engram-* commands. No files to drop by hand.

Update: opencode plugin engram-opencode -f. Uninstall: remove engram-opencode from the plugin array in your opencode config.

./install.sh        # macOS / Linux
.\install.ps1       # Windows (PowerShell)

The store is shared with the Claude Code & Codex versions (~/.engram/general.redb + per-project <project>/.engram/engram.redb). Only the consolidation ledger (watermark / pending / transcripts) is per-CLI under ~/.engram/opencode/, so the adapters never fight over progress.

These run as the injected, pre-permissioned engram query agent, which runs the bundled engram binary and shows the output verbatim.

Same engine as engram — tiered, ACT-R-style activation, self-forgetting (forgetting = demotion, not deletion).

Runs as an isolated opencode session (created in-process via the SDK client, fired with promptAsync), so its context is independent and the user's session is untouched. Live reviewer sessions are attached under the source session when possible, marked with metadata.engram=true, metadata.engramRole="reviewer", metadata.background=true, and metadata.hidden=true, then archived immediately so opencode desktop does not show the temporary reviewer in normal session lists. It runs as the self-provisioned engram-reviewer agent, pre-permissioned for external_directory + bash (it reads the transcript slice / SKILL.md outside the cwd and runs the engram binary) while edit/webfetch stay denied — so the background session never stalls on a permission prompt. It ends with engram consolidate-done (advances the watermark, clears the pending). If it never finishes (crash / restart), catch-up replays the pending on the next startup — nothing lost, nothing redone.

opencode-plugin/                   the npm package (name: engram-opencode)
  package.json                     main -> plugin/engram.ts
  plugin/engram.ts                 the adapter: injection + idle review + config-hook self-provision
  scripts/reviewer-prompt.md       reviewer prompt (kernel asset, identical across adapters)
  skills/engram/SKILL.md           engram agent interface + judgment rubric (kernel asset)
  bin/                             four-platform engine binaries
install.{ps1,sh}                   local/offline installer (auto-load layout)

The reviewer/query agents and commands are not files — they are injected at runtime by the config hook, so they travel with the package automatically.

• Plugin, not shell hooks: a single engram.ts loaded into opencode's (Bun) runtime.
• session.idle instead of SessionEnd: review runs only when the increment ≥ ENGRAM_REVIEW_MIN_LINES (default 10 — lower than the others' 40 because this adapter normalizes one-message-per-line, not one-event-per-line).
• Transcript via the SDK: client.session.messages(...) normalized to a stable JSONL fed to the same review-prepare (kernel unchanged).
• Reviewer via the in-process SDK (an isolated session as the engram-reviewer agent) rather than a headless claude -p / codex exec subprocess.
• Self-provisioning: agents + commands injected via the config hook instead of shipped as files.

• ENGRAM_REVIEW_MIN_LINES — increment (in normalized transcript lines) needed to trigger a review (default 10).
• ENGRAM_REVIEWER_MODEL — providerID/modelID override for the reviewer session (default: opencode's default model).
• ENGRAM_BIN — absolute path to an engram binary, overriding the bundled one.
• ENGRAM_REVIEWER=1 — makes the plugin inert (used by the reviewer subprocess / for debugging).

1. Install and open opencode in a project; send a message.
2. Injection: ask what do you remember about me? — the answer should reflect your hot-index memories.
3. Commands: run /engram-status — you should see the memory overview.
4. Review: after a substantive exchange (≥ ENGRAM_REVIEW_MIN_LINES lines), the background reviewer is marked and archived by default; verify it by ~/.engram/opencode/watermark.json advancing and pending clearing. For debugging, search archived sessions by metadata.engramRole="reviewer" or title engram-review.

Apache License 2.0 — see LICENSE.