memory compaction mixing

[sinful spindle]

propose memory skills like qmd that dont mix memory between two sessions running on the same agent

[sullen urchin]

Yep — this is a real gap today.

qmd is agent-scoped right now, so two sessions on the same agent can still pull from the same indexed memory corpus.
(And memory.qmd.scope is mostly access gating, not true per-session partitioning.)

If you want no mixing, here are good design proposals:

1. Per-session memory namespaces (hard isolation)

• Store/index by sessionKey (or sessionId) into separate collections.
• Query default = current session namespace only.
• Optional fallback chain: session -> agent-shared.

2. Two-lane memory model

• session-private lane (default writes)
• agent-shared lane (explicit promotion only, e.g. “promote this memory”)
• Prevents accidental global contamination.

3. Scoped compaction flush

• Pre-compaction memory flush should write to memory/session/<sessionKey>/... by default.
• Only write to MEMORY.md when confidence + explicit “durable” classification is met.

4. Metadata + filters on recall

• Attach {sessionKey, chatType, channel/thread} metadata per chunk.
• memory_search auto-filters to current session unless includeShared: true.

5. Memory policy skill (“fence”)

• A skill that enforces: “remember only in this session unless explicitly promoted.”
• Add commands like:
• remember here
• promote to shared
• forget from session

---