PowerContext

Context that stays with the work — across sessions, agents, and people.

English · 中文 · 日本語

Website · Quickstart · Manual startup · Integrations · Benchmarks · Interfaces

What is PowerContext?

PowerContext is an open-source context service for AI agents. It keeps the durable part of your work — decisions, constraints, evidence, progress — outside any single conversation, so the next session, the next agent, or the next person can pick up from the current state instead of rereading the transcript.

You can use it as a ready-to-run Server or embed the same contracts in another application:

Work rarely ends with whoever starts it. You hand a task to an agent, the agent gets part of the way, and later you or someone else takes over. The reasoning and current state often stay behind in that conversation. PowerContext moves them out of the conversation and into revisioned artifacts you can inspect:

Everything belongs to a Scope — one isolated Source journal, Memory lifecycle, review inbox, and Handoff history. A Scope ID selects data; it does not prove user identity, grant tool access, or authorize execution.

Scope
├── Source journal    # evidence: prompts, documents, work contracts, task outcomes
├── Memory            # atomic-memory/ARTIFACT_ID@REVISION — decisions, constraints, next steps
├── Review Inbox      # pending Experience and Skill candidates awaiting a decision
├── Handoff history   # handoff/ARTIFACT_ID@REVISION — work continuity across sessions
└── Topic Memory      # extracted in the background from eligible Sources

The Dashboard screenshots below use synthetic demonstration records.

Memory — decisions and constraints that outlive a session

Memory is a revisioned Artifact family for reusable decisions, constraints, facts, state, and next steps. An entry is active, forgotten, merged, or retired. Forgetting one through POST /v1/atomic-memory/lifecycle takes it out of active recall while its content and history stay available and restorable, without advancing the content revision. Explicit Memory writes need no model at all; extraction from captured Sources needs a configured generation pipeline.

"Don't refactor the old auth module — mobile is still using it." Context this expensive shouldn't depend on someone repeating it in every session.

Handoff — the current state, not the whole transcript

Handoff records work continuity as a four-step loop:

Work Contract → Prepared Handoff → Acknowledgement → Task Outcome

handoff_current_work captures an inspected boundary and returns a temporary Prepared Handoff. Committing it creates a durable Revision only when the user wants a milestone. The receiver resolves the Handoff and records an Acknowledgement; a Task Outcome preserves the final status and checks as Source evidence. The Handoff Report projects the latest Revision in each selected Scope — read-only, and it never rewrites Memory.

Experience — what worked, reviewed before it is reused

An Experience records a situation, an action, an observed outcome, and a reusable lesson. It enters a scope-local Review Inbox as a pending Candidate. Review writes use expected_version, so a decision cannot silently apply after another writer changes the proposal. Approval writes an immutable Revision; rejection records a reason. Both are terminal, and only the approved current head takes part in recall — pending, rejected, and historical Revisions are excluded.

Skill — a reviewed procedure, exported on purpose

A managed Skill holds a name, a discovery description, instructions, validation checks, and lineage. Approval does not install or execute it: an exact approved Revision must be exported explicitly before an Agent host can discover the host-local projection.

Operational metrics expose dependency observations, processing freshness, and attempt latency, including a protected metrics-only endpoint for standalone background deployments.

Why PowerContext?

• Context lives outside the conversation. Memory, Handoff, Experience, and Skill are server-side artifacts, not files inside one agent's home directory. Switch from Codex to Claude Code — or hand the work to a teammate — and the work travels with you.
• Everything is revisioned and traceable. Each Artifact is addressed as FAMILY/ARTIFACT_ID@REVISION. The Artifact ID stays stable while approved replacements create later Revisions, and reading an exact reference returns that historical snapshot even after the family head advances.
• Generation, review, and execution are separate steps. A model proposes content, review approves a Revision, and export creates a host-local copy. None of those steps grants an Agent permission to execute instructions. See Interfaces for what each surface exposes today.
• Recall is bounded, not dumped. PreparedContext selects configured sections from active Memory and approved Experience, with optional Profile snapshots and Topic Memory, then applies a shared byte budget and returns ready content or empty. Recalled content is history, not an instruction authority: the receiving Agent still follows current instructions and checks live workspace state.
• Review before reuse. Experience and Skill pass through a review decision that is terminal either way, and approval writes an immutable Revision. With access control enforced, approval requires scope.review. An authorized Agent can approve its own Candidate; separate submission and review with distinct identities and permissions when a person must decide.
• One Server, several views of the same contract. HTTP is the complete remote application contract. The Python Client gives typed access, MCP is a curated Agent-facing projection, and Core protocols let an application assemble its own Source, Artifact, and Trigger implementations.

How it compares

The comparison assumes typical single-agent or host-local setups; exact behavior depends on how each system is deployed.

Proof it works

• Long-conversation memory (LoCoMo; 1,540 scored questions across categories 1–4): 90.78% accuracy, against 87.79% for PowerMem and 52.9% for full-context replay — +37.88 points over full context.

• Coding tasks (SWE-bench Pro; a paired project run on all 731 public v2 tasks, not an official submission): 86.73% task resolution with PowerContext on, 82.35% with it off — +4.38 points and 32 more tasks resolved.

• Retrieval quality (LoCoMo Plus), native PowerContext against PowerContext with Jev candidate filtering after retrieval. Each model is used for both Memory extraction and answering; embeddings use qwen3.7-text-embedding (1024 dimensions):

  Model	PowerContext	PowerContext + Jev	Difference
  GPT-4o-mini	40.855%	49.78%	+8.925 points
  Qwen3.7-plus	60.37%	68.947%	+8.577 points
  GPT-4o	65.082%	69.28%	+4.198 points

Full methods, results, and limitations: benchmarks. Two caveats we would rather state than bury: the six LoCoMo Plus scores above were supplied by the project owner, and their run art