Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
Python agentaianthropicclaude-codecompressioncontext-engineeringcontext-windowcursorfastapilangchainllmmcpopenaiprompt-engineeringproxypythonragtoken-optimizationtokenstypescript
fix(memory): require explicit updates for supersession (#2188)
Chester6d897e8eaaba
diff --git a/headroom/memory/mcp_server.py b/headroom/memory/mcp_server.pyindex e71fe9b54..52830ff06 100644--- a/headroom/memory/mcp_server.py+++ b/headroom/memory/mcp_server.py@@ -76,7 +76,9 @@ _TOOLS = [description=("Save information to persistent memory for future sessions. ""Use this for decisions, conventions, architecture context, "- "user preferences, project facts, or anything worth remembering.\n\n"+ "user preferences, project facts, or anything worth remembering. "+ "Saving a similar fact does not replace an existing memory; corrections "+ "must use an explicit update path with the existing memory ID.\n\n""IMPORTANT: Break information into atomic facts — one fact per ""entry in the 'facts' array. Each fact should be a single, ""self-contained statement that answers one question. "@@ -264,11 +266,6 @@ async def _handle_search(return [TextContent(type="text", text=f"Search error: {e}")]-# Similarity threshold for auto-supersession: if a new memory is this-# similar to an existing one, it replaces (supersedes) the old one.-_SUPERSEDE_SIMILARITY = 0.70--async def _handle_save(backend: LocalBackend, arguments: dict[str, Any], user_id: str) -> list[TextContent]:@@ -294,44 +291,16 @@ async def _handle_save(if not fact:continue- # Check for semantically similar existing memory to auto-supersede- superseded_id: str | None = None- try:- existing = await backend.search_memories(- query=fact,- user_id=user_id,- top_k=3,- )- for r in existing:- if getattr(r.memory, "superseded_by", None):- continue- if r.score >= _SUPERSEDE_SIMILARITY:- superseded_id = r.memory.id- logger.info(- f"Memory MCP: auto-superseding [{r.memory.id[:8]}] "- f"(similarity={r.score:.2f}): {r.memory.content[:60]}"- )- break- except Exception:- pass-- if superseded_id:- memory = await backend.update_memory(- memory_id=superseded_id,- new_content=fact,- )- results_lines.append(- f" updated [{superseded_id[:8]}→{memory.id[:8]}]: {fact[:60]}"- )- superseded += 1- else:- memory = await backend.save_memory(- content=fact,- user_id=user_id,- importance=importance,- )- results_lines.append(f" saved [{memory.id[:8]}]: {fact[:60]}")- saved += 1+ # Similarity is a retrieval signal, not proof that two memories+ # represent versions of the same fact. Only explicit update paths+ # with a caller-supplied memory ID may create a supersession chain.+ memory = await backend.save_memory(+ content=fact,+ user_id=user_id,+ importance=importance,+ )+ results_lines.append(f" saved [{memory.id[:8]}]: {fact[:60]}")+ saved += 1summary = f"Saved {saved} new, updated {superseded} existing ({saved + superseded} total)"return [TextContent(type="text", text=summary + "\n" + "\n".join(results_lines))]diff --git a/tests/test_memory/test_mcp_server.py b/tests/test_memory/test_mcp_server.pyindex 34339c31d..b86e15179 100644--- a/tests/test_memory/test_mcp_server.py+++ b/tests/test_memory/test_mcp_server.py@@ -228,3 +228,30 @@ def test_search_fails_open_when_access_tracking_fails() -> None:result = asyncio.run(mcp_server_mod._handle_search(backend, {"query": "preference"}, "alice"))assert "Useful preference" in result[0].kwargs["text"]+++def test_save_does_not_supersede_a_semantically_similar_memory() -> None:+ existing = Memory(content="Use compact output by default", user_id="alice")+ saved = Memory(content="Keep semantic compression disabled", user_id="alice")+ backend = SimpleNamespace(+ search_memories=AsyncMock(return_value=[SimpleNamespace(memory=existing, score=0.91)]),+ update_memory=AsyncMock(),+ save_memory=AsyncMock(return_value=saved),+ )++ result = asyncio.run(+ mcp_server_mod._handle_save(+ backend,+ {"facts": ["Keep semantic compression disabled"], "importance": 0.9},+ "alice",+ )+ )++ backend.search_memories.assert_not_awaited()+ backend.update_memory.assert_not_awaited()+ backend.save_memory.assert_awaited_once_with(+ content="Keep semantic compression disabled",+ user_id="alice",+ importance=0.9,+ )+ assert result[0].kwargs["text"].startswith("Saved 1 new, updated 0 existing")