contextkit

Releases

Changelog

Every release of the library and the cloud SDK, newest first.

This list is read from PyPI on every build, so a release shows up here without anyone editing the site. Install a specific version with pip install outhad-contextkitai==VERSION.

1.10.0 Cloud SDK

  • DEFAULT_HOST is now https://api.contextkit.outhad.com, and dashboard links point at https://app.contextkit.outhad.com. https://contextkit.outhad.com is this site. Override the host with host= or OUTHAD_CONTEXTKIT_HOST as before.
  • The localhost fallback is opt-in: allow_dev_fallback now defaults to False. Before, a connection error against the default host quietly switched the client to http://localhost:8000 for the rest of the session. Pass allow_dev_fallback=True, or host=DEV_HOST, for local development.
  • Fixed: the User-Agent header reports the installed SDK version.
  • Fixed: the PyPI project links point at pages that exist.

0.1.4 Library

Security fix: upgrade if you use the context graph. In 0.1.3 and earlier, with the context graph on, search(user_id="alice") could return another user's memories: topic links joined memories of different users, and graph search did not check whose memory it reached. Graph search now stays within the user_id, agent_id, and run_id you pass, and new links only join memories of the same user and agent. A graph built before 0.1.4 can still hold links between users; search ignores them, and reset_context_graph() followed by backfill_context_graph() removes them.

  • State store (state_store): keep the context graph, personalized-retrieval feedback, and cold storage in Postgres, so several processes share them and a restart loses nothing. Install with pip install "outhad-contextkitai[postgres]".
  • Episodic memory (tcmgm), available on request: build_timeline, get_timeline, and get_causal_chain on Memory and AsyncMemory, with settings for storing events on every add() and for how cause → effect links are found.
  • AsyncMemory.search takes use_tcmgm, time_window, and include_causal, like Memory.search.
  • Fixed: cause → effect links between events were never stored, so causal chains came back empty.
  • Fixed: AsyncMemory.add(..., infer=False) raised AttributeError.

0.1.3 Library

Four opt-in layers, each off by default. With them off, Memory behaves as it did in 0.1.2.

  • Context graph (context_graph): memories linked by replies, timeline order, shared documents, and topic similarity, with graph-first search (BFS or PageRank), decay, optional LLM-labelled edges, snapshot and replay, and a change feed you can send to a webhook, SSE, or Kafka.
  • Personalized retrieval (mspr): feedback, usage, past-success, and intent signals in the ranking, and role policies.
  • Tenants (tenant): a registry of tenants, sub-tenants, and roles, with export, import, migrate, and delete.
  • Decay and versioning (decay_v2): decay scores, archiving, a background scheduler, reference tracking, and cold storage on disk or S3.

Also:

  • history_db_path accepts a Postgres URL.
  • Graph memory (TCMGM): faster fused search, rules-first causal extraction with an LLM fallback, a circuit breaker that falls back to rules when the LLM fails, and optional larger CLIP / SigLIP image models with fp16 or int8 quantisation. Importing the package no longer loads torch or transformers.
  • Fixed: the anthropic provider now sends only temperature, so Claude 4.x models (Haiku 4.5, Sonnet 4.6, Opus 4.6) work. See Providers.
  • Fixed: adding images with vision on keeps the JSON-mode contract.

1.9.0 Cloud SDK

Adds observability: usage, requests, top_users, and graph_snapshot.

0.1.2 Library

  • Apache 2.0 LICENSE file added to the package.
  • Setup guides: a quick start with no infrastructure, Neo4j setup (Desktop, Docker, Aura), and a troubleshooting guide covering installation, imports, connections, API keys, memory operations, TCMGM, and performance.
  • Fixed the author email format in the package metadata.