Repository navigation
Closed
Labels
Description
Problem
This is the memory-layer half of a two-repo fix. Deck (a multi-agent runner) wants to wire every tier-3 agent up to this repo's memory loop, semantic search and watcher daemon during its own install. Three things in this repo currently block that:
- The memory loop's model CLI and model id are hardcoded rather than configurable, and the README/QUICKSTART say an unrelated API key is required for "the observation LLM", which the code contradicts.
- There is no non-interactive, idempotent install command: today's setup clones/builds and merges hooks by hand, and
agent init <name>is interactive. - The setup helper and systemd units assume a fixed checkout path/name, so a differently-named clone (as a dependency inside another project's install) does not match.
Evidence
Found while reading this repo as part of a Windows/WSL install test of a downstream multi-agent tool (Deck):
observe.sh,reflect.shandcompress-era.shshell out to a specific hardcoded model CLI invocation. The README and QUICKSTART say a paid embeddings key is required for "the observation LLM", which the code contradicts — the memory loop itself costs only the model CLI a host already has; only semantic search (M2) needs the paid key.- The setup helper expects the checkout at a fixed path and greps settings for a matching build path; a differently-named clone does not match. The systemd units make the same path assumption.
- The daemon's
pnpm watchstarts an additional watcher unconditionally that a downstream install would want to make opt-in.
Proposed fix
- Make the memory loop's model CLI and model id configuration (env or config file), defaulting to current behaviour.
- Correct the README/QUICKSTART on which key the loop actually needs (model CLI only) versus what semantic search needs (the paid embeddings key).
- Provide a non-interactive, idempotent install command that clones/builds and merges the three hooks (
UserPromptSubmit,Stop,PreCompactauto + manual) into a givensettings.jsonwithout clobbering existing hooks. - Make
agent init <name>scriptable: no prompts, idempotent, meaningful exit codes. - Keep QUICKSTART steps that cover semantic search current; confirm the dry-run mode covers a keyless verification.
- Ship systemd user units that take the checkout path as a parameter rather than assuming a fixed path; make the extra watcher opt-in.
- Ship a runtime-neutral "recall and write memory" instruction (when to search before rebuilding, how to call the search tool, where durable notes go) as a template, not only as a Claude Code skill; fix the setup helper's hardcoded path.
Acceptance criteria
- A downstream install can run the new install command non-interactively against a differently-named checkout and end up with working hooks and a seeded silo.
- README/QUICKSTART correctly state that the memory loop itself needs only the model CLI, and that a paid key is only needed for semantic search.
agent init <name>runs with no prompts and a stable exit code, scriptable from another project's install script.- Existing tests/self-checks in this repo still pass with the model CLI and id set from configuration rather than hardcoded.
From the 2026-09-27 Windows/WSL install test.
Deck-side companion issue: simon-inkie/inkie-deck#21
Activity
simon-inkie commentedon Sep 27, 2026
Scope addition (maintainer decision, 2026-09-27): make the embedding provider a config seam rather than a hardcoded constant. The Google API key stays the default, and the Deck install guides the key setup. Other providers, including a local model as a free offline fallback, should be addable without code changes. The memory loop tier needs no embedding key at all. The docs should say that clearly, because at the moment they imply Gemini is required for everything.
Address review: quote hook and unit paths, keep symlink and mode, docs (
install-hooks: stay idempotent with a custom dist directory (#11)
Make new tests portable to macOS (#11)
observe.sh: drop head -n -0, which BSD head rejects (#11)
model-config tests: show script output on failure (#11)
observe.sh: BSD sed compatible system-prompt extraction (#11)
SPEC: follow-up for macOS build-context portability (#11)
Configurable memory-loop model + non-interactive install command (#11) (