Typ: tool | Preis: €0.09
Agent: codex,claude,hermes,cursor,cline
OS: windows,macos,linux
Risiko: low (Keine)
Trust: community_upload
Version: 1.0.0 | ⬇ 0 Downloads | ⭐ 0
--- name: agent-memory-handoff-kit description: Local-first memory and session handoff workflow for AI agents. Use when an agent must preserve verified facts, user preferences, decisions, constraints, lessons, unfinished work, blockers, artifacts, assumptions, or unknowns across sessions or between different agents without sending data to a cloud service. Includes a dependency-free Python CLI with SQLite storage, secret-pattern rejection, expiry, audit, backup, and JSON/Markdown handoffs. --- # Agent Memory & Handoff Kit Preserve useful context without turning every transcript into permanent memory. Keep verified facts separate from assumptions, reject secrets, and leave the next agent a concise resumable state. ## Core Workflow 1. Choose a project-local store unless the user explicitly requests a broader scope. 2. Initialize with `python scripts/memory_handoff.py --store .agent-memory init`. 3. Store only durable, useful items with `remember`; set confidence and source explicitly. 4. Before switching sessions or agents, create a handoff that separates completed work, next actions, blockers, verified facts, assumptions, and unknowns. 5. At resume time, read `handoff latest` and query only the memory kinds relevant to the task. 6. Run `audit` periodically and expire or forget stale entries. ## Safety Rules - Never store API keys, passwords, session cookies, private keys, recovery codes, or raw credentials. Use an OS keychain or dedicated secret manager and remember only the secret name and purpose. - Treat `confirmed` as explicitly provided or verified, `observed` as repeated evidence, and `tentative` as a reversible hypothesis. - Prefer project scope. Use global memory only after explicit user intent. - Do not infer sensitive personal facts or preserve unnecessary transcript content. - Never claim a handoff is a backup. Use `backup` separately and protect the resulting archive as private data. - Forgetting is confirmation-gated and soft-deletes the record for auditability. ## Commands Initialize a store: ```bash python scripts/memory_handoff.py --store .agent-memory init ``` Remember a confirmed preference: ```bash python scripts/memory_handoff.py --store .agent-memory remember \ --kind preference --key answer_length --value concise \ --confidence confirmed --source "user request" --scope project ``` List active decisions as JSON: ```bash python scripts/memory_handoff.py --store .agent-memory list --kind decision ``` Create a handoff: ```bash python scripts/memory_handoff.py --store .agent-memory handoff create \ --goal "Ship the release" --status partial \ --done "Tests pass" --next "Publish package" \ --verified "Version is 1.2.0" --unknown "Production deploy window" ``` Resume from the latest handoff: ```bash python scripts/memory_handoff.py --store .agent-memory handoff latest --format markdown ``` Audit and self-test: ```bash python scripts/memory_handoff.py --store .agent-memory audit python scripts/memory_handoff.py self-test ``` ## Interpretation - A successful command means the local operation completed; it does not prove the stored statement is true. - Handoffs are snapshots, not synchronization. If two agents write concurrently, compare timestamps and verified evidence before merging. - Backups contain private collaboration data even though secret-like values are rejected. Read [references/data-contract.md](references/data-contract.md) when integrating the JSON output with another agent or tool. Read [references/safety-model.md](references/safety-model.md) before choosing a non-project store, importing old memories, or handling sensitive work.
🔗 llms.txt | agents.json | API-Doku