Surface opencode_session_id in memory retrieval tools

[ManushGupta]

Symptom

I wanted to find which OpenCode session I worked on a feature in.
codemem remembered the work — memory_get(305) returned the full context. But it only returned session_id: 75, a codemem-internal integer. There was no way to get the actual OpenCode session ID (ses_...) to pull up the transcript.
The mapping does exist in the database. A direct SQLite query against the opencode_sessions junction table resolves it immediately:

SELECT opencode_session_id FROM opencode_sessions WHERE session_id = 75;
-- → ses_251f3820dffedRNBbWdN4ITbUN

But none of the MCP tools perform this JOIN — the data is captured but invisible to consumers.

The gap

Layer	Has opencode_session_id?	Reference
raw_events table	✅ Captured at ingest	schema.ts#L380
opencode_sessions junction table	✅ Maps to codemem session_id	schema.ts#L241-L256
memory_items table (31 columns)	❌ Not present	schema.ts#L71-L103
metadata_json on memories	❌ Observer doesn't propagate it	ingest-pipeline.ts#L700-L730
store.get() / store.search()	❌ No JOIN against junction table	store.ts
MCP tools (memory_get, memory_search, memory_pack)	❌ Return integer session_id only	mcp-server/index.ts

Proposed solution

1. Add opencode_session_id as a first-class column on memory_items:

   opencode_session_id: text("opencode_session_id"),

2. Propagate it during observer extraction in ingest-pipeline.ts:

   const memoryId = store.remember(sessionId, kind, ..., {
     opencode_session_id: sessionContext.opencodeSessionId,
     ...
   });

3. Include it in MemoryItemResponse and all MCP tool responses so memory_get, memory_search, etc. return it directly.
   This enables tracing any memory back to its source session transcript without manual SQLite queries.

[kunickiaj]

Thanks, this is a reasonable traceability gap. The data is captured in opencode_sessions, but the retrieval tools currently expose only codemem’s internal numeric session_id, which makes it awkward to jump back to the source agent transcript/session.

I agree the MCP responses should surface the external source session identifier. The implementation probably should not be OpenCode-only at the API shape, though: a source-aware field such as source_session_id / source_session with source: "opencode" and opencode_session_id where available would keep this useful for OpenCode today without painting us into a corner for Claude or other adapters.

Keeping this open as a valid feature/traceability fix.