ArgoBrain-Memory

Portable, evidence-linked memory for AI agent sessions.

Argobrain Memory keeps raw session transcripts authoritative and writes all memory as rebuildable derived artifacts:

• append-only compression episodes;
• typed fact ledgers with evidence pointers;
• active derived summaries;
• SQLite FTS recall indexes;
• replay benchmarks for retention across repeated compaction.

This repository is the OSS-clean extraction of a private Cortex/Hermes memory design. It ships with synthetic fixtures only. It does not include private transcripts, operator paths, credentials, host inventory, or live agent hooks.

Why

Naive agent memory often recursively summarizes summaries. That loses small but operationally important details: paths, commands, decisions, blockers, preferences, and routing facts.

Argobrain Memory uses a different contract:

1. Raw transcript/event log is the source of truth.
2. Episodes and typed facts are append-only derived memory.
3. Summaries and indexes are rebuildable views.
4. Every fact points back to evidence.
5. Benchmarks compare memory modes before adoption.

Modes

Use modes to compare behavior:

off       no enhanced memory; raw transcript only
baseline  old-style rolling prose summary
cortex    full typed ledger, episodes, active summary, FTS recall
external  placeholder adapter mode for future GBrain/Mem0/Graphiti comparisons

CLI example:

python -m argobrain_memory.cli benchmark \
  --fixture fixtures/demo-session \
  --mode cortex \
  --output /tmp/argobrain-memory-demo

Run all modes:

python -m argobrain_memory.cli benchmark \
  --fixture fixtures/demo-session \
  --mode all \
  --output /tmp/argobrain-memory-demo

Artifact Layout

memory-root/
  active-derived-summary.md
  facts.jsonl
  recall.sqlite
  episodes/
    <episode-id>.json

Safety

• No hooks run by default.
• No cloud calls are made by this core package.
• No generated summary is doctrine.
• SQLite indexes are rebuildable and non-authoritative.
• Public examples use synthetic data only.

Status

Initial 0.1.0 release. The core architecture is stable enough for benchmarking and adapter development. Public package publishing should still wait for a fresh-clone/fake-HOME release gate.