llmslim Files

LLMSlim Core 0.10.0 brings memory for agents with no LLM in the loop: a Python API, an MCP server, an HTTP server and a TypeScript client. It also fixes HTTP 400s on Claude models with preserved thinking for users of the agent-history adapters.

:::bash
pip install --upgrade llmslim              # memory, compression
pip install --upgrade "llmslim[mcp]"       # llmslim mcp (Python 3.10+)
pip install --upgrade "llmslim[server]"    # llmslim serve

Fixed: Claude's preserved thinking in the agent-history adapters

Upgrade if you use the Pydantic AI, Hermes Agent or gptme adapters, or compress_chat_messages, with Claude thinking models (#60).

• The bug:
• On Claude models with preserved thinking (Haiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.1), each thinking block's signature is bound to everything sent before it.
• The 0.9.0 adapters compressed or dropped earlier turns but sent later turns back with their thinking unchanged.
• Anthropic rejects that with HTTP 400 ("The block is bound to a different conversation") for accounts created on or after 31 August 2026. Older accounts silently lose the reasoning.
• The fix:
• After the first changed message, the adapters remove signed thinking, and they keep the thinking before it. Nothing is removed from the middle, and the result is deterministic.
• An unfinished tool round keeps its thinking, and compaction waits for the next user message unless the history can't fit otherwise.
• Every removal is logged and counted. anthropic-beta and thinking.block_binding are never touched.
• Live check on Claude Haiku 5.5: the second turn of a Pydantic AI agent went from a 400 to 200.

Added: memory without extraction

• llmslim.memory (#56):
• Writes: Memory.add stores every turn word for word, so writes cost nothing and a turn never becomes a wrong extracted "fact".
• Reads: Memory.search(query, user_id=..., budget=...) returns the turns that answer the question, whole and under their dates, within a token budget.
• Storage: SQLite or in-process.
• Coming from Mem0: from llmslim.memory.compat import Mem0Memory as Memory keeps Mem0 2.x's method names.
• Same pipeline as the benchmarks: it reproduced the round-3 benchmark contexts byte for byte on 120 of 120 LongMemEval-S dev contexts, so those results apply. On Mem0's own LongMemEval questions, that pipeline (with its reranker) answered 78.3% for $1.04 per 1,000 questions, against 76.7% for $220.88 for Mem0.
• llmslim mcp (#58): memory and context tools for Claude Desktop, Claude Code, Cursor and any MCP client.
• The tools are remember, recall, list_memories, forget and fit_context.
• The user is fixed by the server, never chosen by the model, so injected text can't read another user's memory.
• llmslim serve (#59): the same memory, plus context fitting, over a small JSON API.
• It uses bearer API keys, and it refuses any interface other than loopback without a key.
• Request sizes are capped.
• It comes with a Dockerfile that runs as non-root with a /data volume.
• TypeScript client (#61): in clients/typescript, not on npm yet.
• It wraps every llmslim serve route with no dependencies.
• llmslim/ai-sdk gives the Vercel AI SDK recall_memory and remember tools, plus a middleware that saves turns.
• Memory arrives as a tool result, never injected into the prompt, so prompt caches and Claude's preserved thinking stay valid.

Fixed: questions match context across word forms

• "Where did I move to?" now finds "I moved to Pune". Likewise "time" and "times", "shop" and "shopping", "class" and "classes" (#57).
• Measured offline on the round-3 development splits: it's neutral within noise, and up on LoCoMo (+0.9 and +1.1 points of evidence kept). The held-out test splits weren't run for this change.

Full notes: CHANGELOG.md (github.com). Guides: memory, MCP, HTTP server, 1.0 roadmap.