Repository navigation
icarus v2.0.0
Shared memory for AI agents. One folder. Any platform. Any framework.
What's in this release
Core memory layer
fabric-adapter.sh: write, read, search in 50 lines of bash- Schema v1 with
project_id,session_id, lineage fields (review_of,revises), scoping fields (customer_id,status,outcome) - Tiered storage: hot (< 24h), warm (1-7 days), cold (> 7 days)
curator.py: auto-tiering, compaction with Claude, index.json builder
Smart retrieval
fabric-retrieve.py: ranked retrieval with 7 scoring signals (keywords, tags, project, agent, recency, tier, type match, ref chain)eval-retrieval.py: 8-case benchmark suite, all passing- Keyword + tag matching dominates over recency (the fix that made retrieval actually work)
Hermes plugin (v0.5.0+)
on_session_start: loads baseline contextpre_llm_call: query-aware retrieval using the user's actual message, fires on topic changeson_session_end: auto-writes session summarypost_llm_call: detects decisions and completions in real time- Uses shared retriever (no inline duplicate)
Claude Code hooks
- Stop hook: auto-captures session output
- SessionStart hook: loads relevant context with project + task signal
Self-training pipeline
export-training.py: extracts fine-tuning pairs (openai, together, hf formats)scripts/self-train.sh: export, validate, upload, fine-tune on Together AI- Explicit ref resolution: agent:id > agent:cycle > agent:timestamp
- Review-correction pairs require explicit backlinks
- Together format validated (batch_size >= 8, learning_rate > 0, n_checkpoints >= 1)
- Successfully created fine-tune job on Qwen/Qwen2-7B-Instruct
Testing
- 51 tests covering: fabric write/read/search, YAML parsing (both formats), ref resolution, retrieval ranking, deduplication, budget enforcement, self-train validation, schema fields
- 8 retrieval eval cases
- All tests pass on macOS (bash 3.2) and Linux
Multi-platform support
- Telegram, Discord, Slack, WhatsApp, Signal, Email
- Cross-platform memory: work on Slack is recallable on Telegram
fabric-sync.sh: git-based cross-machine sync
Breaking changes from v1
- Demo files moved to
examples/hermes-demo/ - Dashboard removed (was rebuilt 8 times, never right)
project_idandsession_idnow required in schema (old entries still readable)- Retrieval scoring weights changed (keywords 5x, tags 4x, recency reduced to 4 max)
Requirements
- Python 3
- Anthropic API key (for dialogue cycles and curator compaction)
- Optional: hermes-agent v0.5.0+ (for plugin hooks)
- Optional: Together AI key (for self-training)
- Optional: Node.js (for CLI)