Shared memory and handofffor your coding agents.
PowerContext is a local context service that stores your project's decisions, constraints, verified checks, and task progress. Switch between Claude Code, Codex, or teammates and pick up directly from the latest state. How it works
Hand off this work with PowerContext.
Checks 41 passed · 1 blocked
Saved atomic-memory/m_k2f9@2
Prepared handoff/h_7c3d@3
Handoff
Retry failed webhooks
handoff/h_7c3dhandoff/h_7c3d@3
- Work ContractWaitingObjective and checks agreed
- Prepared HandoffWaitingClaude Code · 41 checks passed, 1 blocked
- AcknowledgementWaitingCodex resolved revision 3
- Task OutcomeWaitingRecorded when the work ends
Travels with the work
- Memory
- 4 active
- Sources
- 3 links
- Next action
- Unblock signature check
Continue this PowerContext Handoff.
Waiting for a Handoff…
Accepted handoff/h_7c3d@3
- Objective
- Retry failed webhooks
- Memories
- 4
- Sources
- 3 links
- Next
- Unblock signature check
One project memory across every session.
When a session ends mid-task, PowerContext keeps the decisions you made, the code state, and the next step ready for whichever agent takes over.
handoff/h_7c3d@3
- Objective
- Retry failed webhooks
- Verified
- 41 checks passed
- Blocked
- Signature check fails on replayed events
- Next action
- Unblock signature check
The next session starts from a structured handoff of the current state.
- Webhook retries back off exponentially, five attempts at most@3activeRead the signing secret exclusively from the vault@1activeRetry queue lives in Redis@2retiredUse idempotency keys on webhook POST@1mergedStaging deploys wait for manual approval@1forgotten
atomic-memory/m_k2f9@3
Webhook retries back off exponentially, five attempts at most
Failed webhook deliveries retry with exponential backoff from 2 s, capped at five attempts. After the fifth failure the event moves to the dead-letter list for review.
Source references
- Claude Code session · retry design
- Pull request #482 review
- Task Outcome · partial
Project decisions and constraints stay saved as versioned revisions.
Experience candidate
- Situation
- Replayed webhooks failed signature checks
- Action
- Verify against the raw body before parsing
- Outcome
- 41 of 41 replays accepted
- Lesson
- Sign and verify raw request bytes directly
Approved as revision 1
Working fixes go through human review before entering shared memory.
replay-failed-webhooks
Replay dead-lettered webhook events safely.
Validation checks
- Idempotency key present
- Signature verified before replay
Approved · revision 2Exported to CodexApproved workflows become reusable Skills exported to your agents.
- Byte budget2.9 of 4 KB
- Memory1.6 KB
- Experience0.9 KB
- Topic Memory0.4 KB
Recall stays within a strict byte budget, injecting only relevant context each turn.
Switch coding agents.Keep the full context.
Start a feature in one coding agent and finish it in another. PowerContext keeps your project's memory, task progress, and approved skills in one shared server.
- CodexPrompt Hook + MCP + Skillofficial
- Claude CodePrompt Hook + MCP + Skillcommunity
- DeepSeek HarnessHTTP plugin + pc_* + /pccommunity
- Hermes AgentMemoryProvider + /pccommunity
- OpenCodeHTTP plugin + pc_*community
- OpenClawMemory plugin + lifecycle hookscommunity
- Pi Coding AgentExtension + pc_* + /pccommunity
- WorkBuddyPrompt Hook + MCP + Skillcommunity
- BubPython Client pluginevaluation
- Pydantic AIToolsetcommunity
- LangChainMiddlewarecommunity
- LangGraphRecall hook + Memory toolscommunity
Connect several agents at once
The setup command installs, configures, and diagnoses each host you name against the same Server.
powercontext setupMeasured on public benchmarks.
Both results come from the published evaluation harness. The benchmarks page details the test methods and exact scope.
- LoCoMo accuracy across 1,540 scored questions
- 90.78%
- Full-context replay scores 52.9% on the same questions.
- fewer input tokens on SWE-bench Pro
- 25.4%
- About 225 million per 100 tasks with PowerContext on, 302 million with it off.
What people say

PowerContext can automatically extract memory and revise topic summaries. Its formal Experience workflow adds a review gate before reusable lessons become available for recall.

A deployment problem I was debugging in Claude Code yesterday, I could pick up again in Codex today. The idea behind this PowerContext plugin is just to my taste: switch tools without starting over from introductions.

PowerContext moves the durable parts out of the conversation and into a context service, so decisions, constraints, evidence and progress survive the session that produced them.

The context receipt in PowerContext v1.5.0 is a genuinely practical feature. When an Agent got something wrong, you used to have to guess: was it never found, found but not passed on, or passed on and ignored? With receipts, at least the first two can be checked.
From install to your first shared memory
Run a local PowerContext server, enable the memory features you want, and connect your coding agents. Works locally on your machine or across devices.
1Install PowerContext
Install the latest stable release with local Server support. The installer sets up Python and uv if needed; follow its printed PATH instructions.
curl -fsSL https://powercontext.oceanbase.io/install.sh | bash2Follow the configuration wizard
Choose your storage, access scenario, memory capabilities, Dashboard, and Agents. It writes configuration files and tailored next steps.
powercontext config init3Start the Server
Run from the directory containing your generated .env and keep this terminal open. Use the Dashboard URL and token provided by the wizard, if enabled.
powercontext server run --env-file .env4Connect your Agent and verify memory
Follow .env.next-steps.md to connect your Agent and select a Scope. Save a decision, then retrieve it with its citation in a new session using the same Scope.

