A memory OS for your agents
Five capabilities that turn a fixed context window into effectively unlimited working memory.
Smart compaction
When a session grows, Context Controller summarizes verbose content in place — tool outputs, logs, transcripts — and replaces it with a dense summary plus a provenance link back to the original. Facts survive; bytes don't. Choose from built-in strategies or write your own.
Compaction policies
- Summarize-tools — collapses tool outputs to ~20% of original tokens
- Rolling digest — keeps a running summary of the oldest N messages
- Semantic dedupe — merges near-duplicate observations before they bloat the window
- Custom policies — hook your own summarizer via the policy API
Priority tiers
- Pinned — never evicted, never compacted (system prompts, user intent)
- High — protected unless the window is critically full
- Normal — standard lifecycle, compacted before eviction
- Low — evicted first, oldest-first, when pressure rises
Priority-based eviction
Every message carries a priority tier. Under memory pressure the engine evicts the lowest-value content first — deterministically, oldest-first within a tier — so degradation is graceful instead of catastrophic. Your agent's goal and guardrails survive every squeeze.
Semantic recall
Evicted doesn't mean forgotten. Everything removed from the window is embedded and indexed in a recall store. The agent can query it semantically — "what did that PDF say about refunds?" — and pull back just the relevant slice for a fraction of the original token cost.
How recall works
- Evicted content is chunked, embedded, and indexed automatically
- Agents query with natural language; top-k relevant chunks return
- Recall calls are metered separately — typically 5–10% of keeping everything in-window
- Full provenance: every recalled chunk cites its source message
Built for teams of agents
- Shared context pools with per-agent views and access scopes
- Handoffs carry compacted state, not duplicated windows
- Supervisor agents can pin shared facts for the whole swarm
- Per-agent budgets prevent one runaway from starving the rest
Multi-agent shared context
Swarm, pipeline, and supervisor architectures share one managed context pool instead of N duplicated windows. Each agent sees its own view; shared facts are pinned once and referenced everywhere. Handoffs become cheap and lossless.
Observability
Every compaction, eviction, and recall is an auditable event with before/after token counts, the policy that fired, and why. Dashboards show usage per session, per agent, and per policy — so you can tune memory like you tune latency.
What you can see
- Token usage over time, per session and per agent
- Compaction events with reclaimed-token accounting
- Eviction log with priority-tier breakdown
- Recall queries, hit rates, and cost savings vs. full context
- Alerts when sessions approach policy thresholds
See the engine in action
The interactive demo runs the actual compaction policy — in your browser, no signup.