Not only memory
PowerContext is the upgraded version of PowerMem and a context runtime for human-agent collaboration. It turns shared work into project context that can be understood, handed off, and continued.
| Capability | Core value |
|---|---|
| Memory extraction and management | Explicitly record decisions, constraints, outcomes, state, and next steps worth reusing over time; with a generation model configured, Memory can also be extracted from Sources. Revisions and retirements preserve history |
| Bounded request-time recall | Before an agent handles a request, generate one schema-validated, cited PreparedContext based on project scope, relevance, and a byte budget; recall failures do not block the original task |
| Handoff | Organize the objective, verified progress, blockers, next step, and evidence into an inspectable work package so another session, task, model, or agent host can continue from a clear state |
| Sources and evidence lineage | Preserve the original sources of knowledge and link Memory and Artifacts with exact citations; capturing a prompt creates only a Source and does not directly turn it into Memory |
| Experience and Skill governance | A model or caller can only submit a Candidate; an immutable revision is created only after Review, and a Skill must still be exported explicitly—it cannot approve, install, or execute itself |
| Local and service deployment | Use SQLite directly for local development, choose OceanBase for team deployments, and integrate with existing systems through HTTP/OpenAPI, MCP, authentication, and OpenTelemetry |
| Metric | PowerContext | PowerMem | Full-context baseline |
|---|---|---|---|
| Accuracy | 90.78% (1,398/1,540) | 87.79% | 52.9% |
| Search p95 latency | 1.38 s | 1.44 s | 17.12 s |
| Answer tokens per question | ~1.65k | ~0.9k | 26k |
The PowerContext result covers all 1,540 scored questions across the benchmark's 10 conversations. See the LoCoMo evaluation details for dataset selection, retrieval, judging, latency, token, and Artifact boundaries.
In a paired evaluation over all 731 tasks in SWE-bench Pro public v2, enabling PowerContext increased the task resolution rate from 82.35% to 86.73%, an improvement of 4.38 percentage points.
The evaluation ran in a Codex environment, with both the PowerContext OFF and ON groups using the gpt-5.6-sol
model.
| Result | PowerContext OFF | PowerContext ON | Change |
|---|---|---|---|
| Resolved tasks | 602 / 731 | 634 / 731 | +32 |
| Resolution rate | 82.35% | 86.73% | +4.38 percentage points |
SWE-bench Pro public v2 evaluation details.
PowerContext provides official plugins and installation guides for Codex, Claude Code, and DeepSeek Harness. All three integrations use the same scoped data and history-preserving contracts through PowerContext Server; the plugins do not start or embed the Server.
Codex |
Claude Code |
DeepSeek Harness |
You need macOS or Linux, Python 3.11 or newer, uv, and Codex CLI.
uv tool install "powercontext[cli,server]==0.0.2"
# Choose one or more integrations.
powercontext setup codex --source oceanbase/powercontext --ref v0.0.2
powercontext setup claude-code --source oceanbase/powercontext --ref v0.0.2
powercontext setup dsh --source oceanbase/powercontext --ref v0.0.2The first command installs the CLI and local Server in an isolated environment. The subsequent setup commands install the corresponding plugins from the matching repository tag. Run setup again to refresh an existing installation.
Keep the Server running in one terminal:
powercontext server runIn another terminal, verify the service and plugin:
powercontext doctor
powercontext doctor codex # or: claude-code / dshBy default, the Server listens on 127.0.0.1:8000, exposes Streamable HTTP MCP at /mcp, and persists data in a
local SQLite database. Explicit Memory operations work without configuring an inference provider.
Install the locked development environment and hooks:
make installRun the main validation commands before opening a pull request:
make check
make test
make docs-testAfter changing openapi/powercontext.yaml, run make contract-test. See CONTRIBUTING.md for the
complete workflow and docs/en/development/ for implementation guides.
Questions and feedback are welcome in Discord. Use GitHub Issues for reproducible defects and focused feature requests.
PowerContext is licensed under the Apache License 2.0.