Memory DAG Compactor
What it does
Standard memory compaction is lossy — older entries get truncated and details disappear forever. Memory DAG Compactor replaces flat compaction with a directed acyclic graph (DAG) of hierarchical summaries inspired by lossless-claw's Lossless Context Management approach.
Each depth in the DAG uses a purpose-built prompt tuned for that abstraction level:
| Depth | Name | What it preserves | Timeline granularity |
|---|---|---|---|
| d0 | Leaf | File operations, timestamps, specific actions, errors | Hours |
| d1 | Condensed | What changed vs. previous context, decisions made | Sessions |
| d2 | Arc | Goal → outcome → carries forward | Days |
| d3+ | Durable | Long-term context that survives weeks of inactivity | Date ranges |
The raw MEMORY.md entries are never deleted — only organized into a searchable, multi-level summary hierarchy.
When to invoke
- Automatically nightly at 11pm (cron) — compacts the day's memory entries
- When MEMORY.md grows beyond a configurable threshold (default: 200 entries)
- Before a long-running task — ensures memory is compact and searchable
- When the agent reports "I don't remember" for something that should be in memory
How to use
bashpython3 compact.py --compact # Run leaf + condensation passes python3 compact.py --compact --depth 0 # Only leaf summaries (d0) python3 compact.py --compact --depth 2 # Condense up to d2 arcs python3 compact.py --status # Show DAG stats and health python3 compact.py --tree # Print the summary DAG as a tree python3 compact.py --search "deployment issue" # Search across all depths python3 compact.py --inspect <summary-id> # Show a summary with its children python3 compact.py --dissolve <summary-id> # Reverse a condensation python3 compact.py --format json # Machine-readable output
Procedure
Step 1 — Run compaction
bashpython3 compact.py --compact
The compactor:
- Reads all entries from MEMORY.md
- Groups entries into chunks (default: 20 entries per leaf)
- Generates d0 leaf summaries preserving operational detail
- When leaf count exceeds fanout (default: 5), condenses into d1 summaries
- Repeats condensation at each depth until DAG is within budget
- Writes the summary DAG to state
Step 2 — Search memory across depths
bashpython3 compact.py --search "API migration"
Searches raw entries and all summary depths. Results ranked by relevance and depth — deeper summaries (d0) are more detailed, shallower (d3+) give the big picture.
Step 3 — Inspect and repair
bashpython3 compact.py --tree # Visualize the full DAG python3 compact.py --inspect s-003 # Show summary with lineage python3 compact.py --dissolve s-007 # Reverse a bad condensation
Depth-aware prompt design
d0 (Leaf) — Operational detail
Preserves: timestamps, file paths, commands run, error messages, specific values. Drops: conversational filler, repeated attempts, verbose tool output.
d1 (Condensed) — Session context
Preserves: what changed vs. previous state, decisions made and why, blockers encountered. Drops: per-file details, exact timestamps, intermediate steps.
d2 (Arc) — Goal-to-outcome arcs
Preserves: goal definition, final outcome, what carries forward, open questions. Drops: session-level detail, individual decisions, specific tools used.
d3+ (Durable) — Long-term context
Preserves: project identity, architectural decisions, user preferences, recurring patterns. Drops: anything that wouldn't matter after 2 weeks of inactivity.
Configuration
| Parameter | Default | Description |
|---|---|---|
chunk_size | 20 | Entries per leaf summary |
fanout | 5 | Max children before condensation triggers |
max_depth | 4 | Maximum DAG depth |
token_budget | 8000 | Target token count for assembled context |
State
DAG structure, summary content, and lineage stored in ~/.openclaw/skill-state/memory-dag-compactor/state.yaml.
Fields: last_compact_at, dag_nodes, dag_edges, entry_count, compact_history.
Notes
- Never modifies or deletes MEMORY.md — the DAG is an overlay
- Each summary includes a
[Expand for details about: ...]footer listing what was compressed - Dissolve reverses a condensation, restoring child summaries to the active set
- Inspired by lossless-claw's DAG-based summarization hierarchy and depth-aware prompt system
