doobidoo/mcp-memory-service v10.35.0

v10.35.0 — Session-level memory ingestion (LongMemEval R@5 86.0%)

on GitHub

feat: session-level memory ingestion

New memory_store_session MCP tool and POST /api/sessions HTTP endpoint store a full conversation as a single memory unit — all turns concatenated as [role] content lines, stored with memory_type=session and auto-tagged session:<id>.

Why two ingestion strategies?

Turn-level storage (default) is best for precise fact retrieval. Session-level storage improves session-recall benchmarks.

Biggest gains: multi-session +15.2%, temporal-reasoning +10.6% — zero LLM API calls.

What's Included

New MCP Tool — memory_store_session

• Accepts a list of {role, content} turn objects + optional session_id, tags, metadata
• Concatenates all turns as [role] content lines into a single memory body
• Stores with memory_type=session and auto-tag session:<id> (UUID if not provided)
• 10 new handler tests

New HTTP Endpoint — POST /api/sessions

• REST endpoint mirroring the MCP tool for HTTP API consumers
• Same concatenation and tagging semantics
• 7 new endpoint tests

LongMemEval --ingestion-mode flag

• --ingestion-mode session|turn|both for direct apples-to-apples strategy comparison
• Session results: R@5 86.0%, R@10 92.1%

Memory Type Ontology

• session and conversation_turn added as first-class memory types

README: MemPalace comparison

• New section with honest benchmark gap context and explanation of why session-level ingestion narrows the gap to MemPalace-style approaches

Test Count

1,537 tests passing (+17 new: 10 handler + 7 HTTP endpoint)

Upgrade Notes

No breaking changes. Existing turn-level storage and retrieval is unchanged. The new memory_store_session tool is additive.

Full Changelog: https://github.com/doobidoo/mcp-memory-service/blob/main/CHANGELOG.md#10350---2026-04-08