Skip to content

CHANGELOG.md

Latest commit

 

History

History
153 lines (126 loc) · 7.08 KB

CHANGELOG.md

File metadata and controls

153 lines (126 loc) · 7.08 KB
Raw

Changelog

All notable changes to TeleMem are documented here. The format follows Keep a Changelog.

[1.8.0] - 2026-07-11

Fixed

  • Character memory extraction: the per-character prompt received the parse_messages helper function instead of the parsed current dialogue, so character profiles were built from a function repr rather than the conversation. Character extraction now sees the actual turn.
  • add() without a user_id now stores memories in the shared "events" scope (matching add_batch()), so search() — which always includes "events" — can actually retrieve them.
  • search(limit=N) now enforces the limit on the merged result set across scopes (highest score first); previously searching a profile plus "events" without a reranker could return up to 2N results.

Added

  • add(infer=False) and add_batch(infer=False) are now honored: message contents are stored verbatim with zero LLM calls, and the message role is recorded in metadata.
  • add(prompt=...) is now honored as a custom extraction prompt (system prompt override, raw transcript as the user message); also threaded through add_batch().
  • add(memory_type="procedural_memory") now delegates to mem0's procedural-memory pipeline; any other non-None memory_type raises a validation error instead of being silently ignored.
  • add() validates that at least one of user_id/agent_id/run_id is provided, matching the documented (and mem0's) contract.
  • Offline contract test suite (tests/test_contract.py): deterministic fake-LLM/fake-embedder tests pinning the public behavior of add/add_batch/search, character-scope separation, raw ingestion, prompt override, and result limits.
  • Multi-NPC reference example (examples/multi_npc.py): five NPCs live through one scene; each gets a private memory profile plus the shared "events" world-state, then recalls the scene from their own perspective.

Changed

  • Inherited mem0/PostHog telemetry is now disabled by default (import telemem sets MEM0_TELEMETRY=False unless already set); documented in the READMEs and API reference.
  • mem0ai dependency pinned to the tested >=2.0,<2.1 range (TeleMem extends private mem0 internals, so an unbounded range was unsafe); requirements.txt regenerated from the tested environment.
  • Test functions no longer return values (removes pytest warnings).

[1.7.1] - 2026-06-12

Fixed

  • MCP registry name uses the canonical GitHub org casing (io.github.TeleAI-UAGI/telemem) so OIDC namespace authorization and the PyPI ownership marker match.

[1.7.0] - 2026-06-12

Added

  • Published to the official MCP registry as io.github.teleai-uagi/telemem: server.json manifest, PyPI ownership marker, and an OIDC-authenticated publish job in the release workflow.
  • New telemem console script (alias of telemem-mcp) so uvx telemem runs the MCP server with zero install, per the registry convention.
  • Evaluation charter (docs/evaluation.md) aligned with The Benchmark Theatre: baselines before architecture, constant base model, deterministic scoring, audited judges, multi-seed reporting with Wilson intervals, cost/latency as first-class metrics, conflict-of-interest disclosure.
  • LongMemEval harness v2 implementing the charter: --system {telemem, full-context, grep}, --seeds N (mean ± std), per-type Wilson 95% intervals, and --validate-judge adversarial judge auditing. The charter applies to new runs; existing published results are unchanged pending team review.

Changed

  • The MCP SDK (mcp>=1.6.0) is now a core dependency — TeleMem is MCP-ready out of the box; telemem[mcp] remains as a no-op alias.

[1.6.0] - 2026-06-12

First release published to PyPI: pip install telemem.

Added

  • Provider config examples with offline tests: Ollama (fully local stack), DeepSeek, and Moonshot (Kimi) — config/config.<provider>.yaml + tests/test_providers.py.
  • Framework integration examples: examples/langchain_memory.py and examples/llamaindex_memory.py.
  • Documentation site at https://teleai-uagi.github.io/telemem/ (mkdocs-material, auto-deployed via GitHub Pages).
  • LongMemEval evaluation harness: baselines/longmemeval/run_telemem.py (experimental; results pending).

Changed

  • Moved the research/evaluation variant of TeleMemory out of the shipped package: telemem/main.py → baselines/telemem/telemem_legacy.py. It backs the ZH-4O benchmark harness (online_query, offline_build_graph_json) and needs tenacity/pytz, which are not core dependencies. Also fixed the broken eval.py import in the TeleMem baseline.

[1.5.0] - 2026-06-12

Added

  • mem0-compatible return values: Memory.add() / add_batch() now return {"results": [{"id", "memory", "event"}, ...]} and Memory.search() returns {"results": [...]} (previously add returned None and search returned a pre-joined string), making import telemem as mem0 a true drop-in replacement.
  • New telemem[video] extra. The default install is now lightweight; the video pipeline dependencies (opencv, yt-dlp, nano-vectordb, azure-identity) moved out of the core requirements. telemem[all] installs everything.
  • telemem.mm_utils now exposes lazy exports (MMCoreAgent, process_video, init_single_video_db, decode_video_to_frames, extract_choice_from_msg), with a clear install hint when the video extras are missing.
  • GitHub Actions CI (pytest on Python 3.10–3.12 + package build check) and a tag-triggered PyPI release workflow using trusted publishing.
  • CONTRIBUTING.md, CITATION.cff, issue and pull-request templates.

Fixed

  • Mem0ValidationError was raised but never imported (a NameError on invalid messages input).
  • _sync_memory_to_vector_store returned an always-empty list due to a variable typo; stored memories are now reported back to the caller.
  • from telemem.mm_utils import MMCoreAgent (as documented in the README) did not work because mm_utils/__init__.py was empty.
  • Importing telemem no longer mutates sys.path or the root logging level.

[1.4.0] - 2026-06-11

Added

  • Model Context Protocol (MCP) server: telemem-mcp with 8 tools (add_memory, search_memories, get_memories, get_memory, update_memory, delete_memory, delete_all_memories, memory_history) over stdio / SSE / streamable-http (docs/MCP.md).
  • uv-managed environment with committed uv.lock.

[1.3.0] - 2026-01-28

  • Stability improvements; TeleMem Tech Report updated to v4 on arXiv.

[1.2.0] - 2026-01-09

  • Core improvements; MiniMax provider support via OpenAI-compatible API.

[1.1.0] - 2025-12-31

  • Memobase and RAG baselines; evaluation harness improvements.

[1.0.0] - 2025-12-05

  • Initial public release: character-aware long-term memory, LLM-based semantic clustering and deduplication, FAISS + JSON dual storage, batch writing, and the multimodal video pipeline (add_mm / search_mm).