agent-memory

E2E encrypted agent memory on IPFS — portable, decentralized knowledge persistence

Agents (Goose, Copilot, Claude Code, observation loops) generate institutional knowledge — decisions, failures, learnings, context. That knowledge currently dies with the session. agent-memory fixes this by providing encrypted, content-addressed, decentralized persistence that any agent can write to and read from.

Why

• Billing integration has been attempted 3 times. Zero traces of those attempts exist in any memory system.
• Observation loop runs 65 times, sees GitHub state, but cannot see decisions made outside GitHub.
• m4 vault has 523 traces, 485 essences — rich personal memory, but agents can't write to it.
• The pipeline's collect-memory.sh can only surface what's already been captured.

The gap: work happens in sessions. Sessions end. Knowledge evaporates. Next session starts from zero.

How it works

• Agent writes a memory entry (JSON: type, content, tags, timestamp, source)
• Entry is encrypted with AES-256-GCM using a key derived from the agent's secret (HKDF, same pattern as wgmesh)
• Encrypted blob is pinned to IPFS via a local Kubo node or Helia (browser/runtime)
• CID is the address — content-addressed, immutable, decentralizable
• An index (also encrypted) maps tags/types → CIDs so agents can query by topic
• Any agent with the key can decrypt and read. Without the key, it's opaque ciphertext on IPFS.