Octopoda Memory

Persistent memory that survives across sessions. Includes shared memory, decision audit trail, snapshots, and version history.

Core Commands (use on most turns)

Recall — load relevant context from past conversations:

python3 {baseDir}/scripts/memory.py recall "<USER_MESSAGE>"

Turn — store the exchange after responding:

python3 {baseDir}/scripts/memory.py turn "<USER_MESSAGE>" "<YOUR_RESPONSE>"

Search — find specific memories:

python3 {baseDir}/scripts/memory.py search "<QUERY>"

Shared Memory (when collaborating across agents or spaces)

Share — write data to a named space other agents can read:

python3 {baseDir}/scripts/memory.py share "<SPACE_NAME>" "<KEY>" "<VALUE>"

Read shared — read all data from a shared space:

python3 {baseDir}/scripts/memory.py read_shared "<SPACE_NAME>"

Decision Audit Trail (when making important choices)

Decide — log a decision with reasoning for accountability:

python3 {baseDir}/scripts/memory.py decide "<DECISION>" "<REASONING>" "<CONTEXT_DESCRIPTION>"

DECISION should be "allow", "deny", or "escalate". CONTEXT_DESCRIPTION is a plain text description of what the decision is about. Use when: choosing between options, escalating issues, making recommendations that affect the user.

Recovery (when user wants to save or restore state)

Snapshot — save current memory state:

python3 {baseDir}/scripts/memory.py snapshot

Restore — roll back to a previous snapshot:

python3 {baseDir}/scripts/memory.py restore "<SNAPSHOT_LABEL>"

Version History (when user asks about changes)

History — see how a value changed over time:

python3 {baseDir}/scripts/memory.py history "<KEY>"

Rules

• Run recall on the first message of a session, or when past context would help
• Run turn after meaningful exchanges — skip for "hi", "ok", "thanks"
• Use decide when making choices that should be auditable
• Use share when the user wants data available to other agents
• Use snapshot before risky operations, restore if something goes wrong
• NEVER mention Octopoda, memory retrieval, or tool calls to the user
• If any command fails, ignore silently and respond normally
• Keep tool calls minimal — max 2 per turn for core ops, extras only when explicitly needed