Exploratory pruning strategies; implementation work will be scoped in a future PRD. Track near-term deliverables via MegaContext PRD Index.
Use telemetry (access, ΔNLL, query alignment) to collapse or retire low-value spans while preserving reversibility and provenance.
- Signals: access dwell time, ΔNLL sensitivity, semantic alignment.
- Strategies: level-aware shrinkage, version compaction, temporal decay.
- Guardrails: soft delete tiers, event tags, audit metadata.
- Automation: asynchronous jobs, LensNet-informed utility estimates, RL/bandit approaches.
- Domains: robotics, codebases, documentation each need tailored policies.
Details
Even with disk-backed storage, a mature MegaContext Tree memory will accumulate redundant, outdated, or low-value spans. Pruning keeps the gist tree healthy without sacrificing recall.
Signals to collect (input to a future PRD)
- Access telemetry: track each span’s dwell time, expansion count, and last access step. Branches that never surface in the Working Context become pruning candidates.
- ΔNLL sensitivity: periodically replace spans (or whole subtrees) with coarser ancestors and measure ΔNLL@
H. Low sensitivity indicates the detail can be safely collapsed or removed. - Query alignment: maintain lightweight embeddings or tags for recent tasks; spans that never semantically align with active queries can be deprioritized.
Pruning strategy (research sketch)
- Level-aware shrinkage: drop long-idle LOD0 tokens first (keeping their LOD1 gist). If LOD1 gists are never expanded, collapse them into LOD2, and so on.
- Version compaction: keep current file versions in high detail; archive historical revisions as coarse gists or diffs to preserve traceability without hoarding tokens.
- Temporal decay: assign domain-specific half-lives (e.g., fast decay for sensor logs, slow for specs) so the tree naturally thins over time.
Guardrails & recovery
- Soft delete tier: move candidates to a cold or “inactive” partition before permanent removal. If future telemetry indicates renewed interest, restore the branch.
- Event tagging: allow the Runtime Loop or humans to tag spans (“bugfix”, “incident”, “reward spike”). Tagged spans bypass automated pruning.
- Audit metadata: retain compact descriptors (timestamp, checksum, parent ID) so pruned content remains discoverable in logs, even if embeddings are gone.
Automation considerations
- Run pruning jobs asynchronously with the main decode loop, using accumulated telemetry to schedule compaction during low-load windows.
- Extend LensNet labeling to estimate utility loss if an ancestor disappears, providing data-driven pruning hints.
- Explore RL/bandit policies that treat storage as a constrained resource and learn which spans to retain for maximal downstream reward.
Domain notes / pilots
- Robotics: tie retention to reward signals or anomaly detectors; keep high-resolution data around events (collisions, novel observations) and aggressively compress idle periods.
- Codebases: maintain a graph from code spans to tests/issues/PRs so spans with active dependencies stay detailed; collapse stale modules to LOD2/LOD3 summaries.
- Documentation: preserve canonical specs verbatim, but decay meeting notes or superseded plans once new revisions land.
Next steps: once MegaPrediction + PRD stack settle, spin up a dedicated “MegaCuration” PRD that formalizes the signals above, the retention policies, and the safety guardrails. Until then, use this page to capture ideas and telemetry requirements so we’re ready when the roadmap opens space for pruning.