Skip to content

build-memory-with-agents.md

Latest commit

 

History

History
100 lines (74 loc) · 2.7 KB

build-memory-with-agents.md

File metadata and controls

100 lines (74 loc) · 2.7 KB

Build Memory with Agents

This is the first practical Mneme playbook for building a reviewed memory pack with agent help.

Goal

Turn raw OpenClaw memory into a reviewed compiled pack through a repeatable flow:

  1. ingest raw evidence
  2. prepare category bundles
  3. send a bundle to an agent
  4. receive candidate JSON
  5. validate the evidence refs
  6. materialize category outputs
  7. merge category outputs into one reviewed pack

Deterministic pieces

Mneme already provides the deterministic shell:

  • scripts/mneme_ingest_memory.py
  • scripts/mneme_llm_compile.py
  • scripts/mneme_materialize_candidates.py
  • scripts/mneme_llm_roundtrip.py
  • scripts/mneme_merge_pack.py
  • scripts/mneme_runtime_orchestrate.py

Current architecture

The architecture is now split cleanly like this:

  • repo scripts = ingest, prep, validate, materialize, merge, runtime bridge
  • runtime tools = send bundle to agent, receive JSON back

That is deliberate. It avoids pretending a standalone repo script can summon agents without the runtime.

Category-by-category flow

1. Prepare bundles

./scripts/mneme_llm_roundtrip.py \
  --root /path/to/workspace \
  --category projects \
  --json

2. Agent pass

Use the prepared bundle as input to an agent and ask for candidate JSON only.

3. Validate + materialize

./scripts/mneme_llm_compile.py validate \
  --raw /path/to/output/raw \
  --input ./candidate-projects.json

./scripts/mneme_materialize_candidates.py \
  --category projects \
  --input ./candidate-projects.json \
  --out /path/to/output/materialized-projects

4. Merge reviewed categories

./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

Proven categories so far

The agent-connected loop has already been proven for:

  • projects
  • systems
  • decisions
  • incidents

That is the first serious Mneme memory pack.

Proven runtime result

The runtime bridge is also proven now:

  • prepare-task generated a real category task
  • the runtime dispatched it to an agent automatically
  • the agent returned candidate JSON
  • apply-result validated and materialized it
  • the reviewed pack merged successfully

So Mneme has now proven both:

  • category-by-category agent compilation
  • automatic runtime dispatch through the bridge layer

What remains

The next work is not basic capability. It is refinement:

  • multi-category automatic runs
  • result normalization to reduce manual cleanup
  • better merge quality and review summaries
  • a first-class runtime command for full Mneme compile jobs