Skip to content

continuation.md

Latest commit

 

History

History
102 lines (75 loc) · 2.55 KB

continuation.md

File metadata and controls

102 lines (75 loc) · 2.55 KB

Mneme Continuation Guide

If you are resuming Mneme work in a new session, start here.

Current status

Mneme now has:

  • memory health check
  • secret scrubber
  • drift detector
  • raw evidence ingest
  • structured compiler outputs
  • LLM-assisted compile scaffold
  • candidate validation + materialization
  • reviewed-pack merge
  • runtime orchestration bridge
  • a proven automatic runtime dispatch loop

The first serious reviewed pack has already been proven across:

  • projects
  • systems
  • decisions
  • incidents

Stable assumptions

Repo:

  • /path/to/mneme

Workspace root:

  • /path/to/workspace

Do not depend on old temporary output directories still existing on the next day. Regenerate fresh outputs.

How to continue

1. Verify the Mneme baseline

cd /path/to/mneme
./scripts/mneme_run.py --root /path/to/workspace --skip-scrub --json

Expected baseline:

  • memory check ok
  • drift = 0 contradictions / 0 stale

2. Prepare a fresh runtime task

./scripts/mneme_runtime_orchestrate.py prepare-task \
  --root /path/to/workspace \
  --category projects \
  --max-items 25 \
  --raw-out /path/to/output/raw \
  --bundles-out /path/to/output/bundles \
  --materialize-out /path/to/output/materialized-projects

3. Dispatch through the runtime

Use the produced taskPrompt with an OpenClaw agent/sub-agent.

4. Apply the result

./scripts/mneme_runtime_orchestrate.py apply-result \
  --category projects \
  --raw-out /path/to/output/raw \
  --candidate /path/to/candidate-projects.json \
  --materialize-out /path/to/output/materialized-projects

5. Merge reviewed categories when ready

./scripts/mneme_merge_pack.py \
  --inputs /path/to/output/materialized-projects /path/to/output/materialized-systems \
           /path/to/output/materialized-decisions /path/to/output/materialized-incidents \
  --out /path/to/output/reviewed-pack

Highest-value next work

  1. Multi-category automatic run

    • one runtime job that iterates categories and merges at the end

  2. Result normalization

    • reduce manual cleanup when agent output uses slightly different field names/shapes

  3. Review summaries

    • generate a compact summary of what changed in the reviewed pack

  4. Better merge quality

    • dedupe overlapping entries across categories where useful

Important lesson

Pure scripts were good for plumbing and hygiene. Useful memory compilation required an agent/LLM judgment layer. The winning architecture is:

deterministic pipeline + agent judgment + deterministic validation/materialization