changelog

Every release of the open-source packages

Newest first, for actrone-memory in Python and TypeScript and for create-actrone-app. Each entry comes from the package’s own CHANGELOG.md, so this page and the repositories never disagree.

RSS feed

7 releases

1 Oct 2026#ts-0.1.2

actrone-memory 0.1.2

TypeScript

npm GitHub release

Framework compatibility fixes, found by type-checking every adapter against the real framework at the oldest and newest release of its supported range, and fact extraction that works on small local models.

Fixed7

Fact extraction returned nothing on small models. With qwen2.5:3b on Ollama, extraction came back empty on 12 of 15 real exchanges, because the model read the assistant's reply as part of what to mine. Extraction spec 1.1 frames the conversation, says whose facts to record, asks for a JSON-schema structured output and gives two worked examples (one with facts, one with none). On the same model it found 14 of 14 expected facts and stored nothing for small talk. OpenAIFactExtractor falls back to JSON mode, for good, only when a server rejects the schema. EXTRACTION_SPEC_VERSION is now "1.1"; EXTRACTION_RESPONSE_SCHEMA and formatExtractionInput are exported for custom extractors.
Passing a real openai client to OpenAIFactExtractor failed to compile. ChatCompleterLike typed each message's role as a plain string, which the openai SDK's message types reject, so new OpenAIFactExtractor(new OpenAI(...)) only worked with a cast. The roles are now the literal ones the extractor sends, and a docs example type-checks the real client in CI.
npm install actrone-memory failed next to current framework releases. The optional peer ranges stopped below the versions most projects now run, and npm refuses such installs (ERESOLVE) rather than warning. For example, a project on the Vercel AI SDK 7 could not install 0.1.1. The ranges now cover the current majors: ai >=5 <8, @mastra/core >=0.10 <2 and @voltagent/core >=0.1.14 <3.
langgraphMemory().loadMemories() did not type-check as a LangGraph message. Its return type was an interface, which TypeScript does not treat as the index-signature record LangGraph's messagesStateReducer accepts. It is now a type alias with the same shape, so it merges into state.messages without a cast.
Floors that pointed at unrelated packages. npm's genkit 1.0.0 to 1.0.3 and agents 0.0.1 and 0.0.2 were published by other projects before Firebase and Cloudflare took the names. The floors are now genkit >=1.0.4 and agents >=0.0.37, the first real releases.
VoltAgent floor. @voltagent/core below 0.1.14 has no instructions option, which voltagentMemory().withInstructions() fills. The floor is now 0.1.14.
Mastra and LlamaIndex.TS recipes showed calls that do not exist. Mastra's memory goes in agent.generate(input, { context: [{ role: 'system', content }] }), and LlamaIndex.TS chat engines take it as the leading message of chatHistory; chat() has no systemPrompt parameter. The recipes, npx actrone-memory add output and adapter docs now show these calls.

1 Oct 2026#py-0.2.1

actrone-memory 0.2.1

Python

PyPI GitHub release

Framework integration fixes found by running every integration against the real framework at the oldest and newest version we support, fact extraction that works on small local models, plus PyPI listing updates.

Fixed7

Fact extraction returned nothing on small models. With qwen2.5:3b on Ollama, extraction came back empty on 12 of 15 real exchanges, because the model read the assistant's reply as part of what to mine. Extraction spec 1.1 frames the conversation, says whose facts to record, asks for a JSON-schema structured output and gives two worked examples (one with facts, one with none). On the same model it found 14 of 14 expected facts and stored nothing for small talk. EXTRACTION_SPEC_VERSION is now "1.1"; the prompt, EXTRACTION_RESPONSE_SCHEMA and format_extraction_input are exported for custom extractors.
Microsoft Agent Framework. The adapter targeted the pre-1.0 API (ChatAgent, ContextProvider.invoking, Context), which Agent Framework 1.x removed. It now implements the 1.x ContextProvider (before_run adds the recalled context through context.extend_instructions, after_run saves the turn from the response), so Agent(client, context_providers=[memory.as_context_provider()]) works.
LangGraph. ActroneCheckpointer was a plain class, so graph.compile(checkpointer=...) rejected it. It is now a real BaseCheckpointSaver that wraps a saver of your choice (saver=, an in-memory saver by default) for graph state and also records each completed human and AI exchange as memory, once per exchange.
Haystack. ActroneRetriever and ActroneWriter were not registered Haystack components, so Pipeline.add_component refused them. They are now real @component classes with declared inputs and outputs, and support both run and run_async.
LlamaIndex. ActroneLlamaMemory now recalls against the user message a chat engine has just stored when get() is called without input, keeps that message in the history it returns, and implements set/aset and the tokenizer that chat engines expect.
AutoGen. update_context now returns an UpdateContextResult wrapping a MemoryQueryResult, as AssistantAgent requires, and query accepts a string or a MemoryContent.
DSPy. ActroneRM now returns passages with long_text and a passages attribute, the shape dspy.Retrieve reads.

Added1

OpenAIFactExtractor works with local and self-hosted models. It takes base_url (for example http://localhost:11434/v1 for Ollama) or a ready client, so any OpenAI-compatible server can do extraction, as the TypeScript library already allowed. It asks for a JSON schema and falls back to JSON mode, for good, only when a server rejects the schema.

Changed3

Newer framework majors are supported. Haystack 3, Agno 2 and 3, Pydantic AI 2 and Google ADK 2 passed the same checks as the older releases, so the extras now allow them: haystack-ai>=2.0,<4, agno>=1.0,<4, pydantic-ai>=1.32,<3 and google-adk>=1.2,<3.
Supported framework versions now match what was tested. Each extra's lower bound is the oldest release the integration was verified against: crewai>=0.95, autogen-agentchat and autogen-core>=0.4.3, smolagents>=1.5.1, openai-agents>=0.2, pydantic-ai>=1.32, semantic-kernel>=1.16 and google-adk>=1.2. Older releases fail to install or import with current dependencies, or lack the API the integration uses. The dspy extra now installs dspy (the package was renamed from dspy-ai) and allows 3.x: dspy>=2.5,<4.
PyPI listing. Keywords now cover what people search for (agent-memory, long-term-memory, semantic-search, rag, llm, pgvector, agno), and new classifiers state that the package is typed, asyncio-based, built on Pydantic 2, Python 3 only and OS independent.

28 Sept 2026#ts-0.1.1

actrone-memory 0.1.1

TypeScript

npm GitHub release

A metadata-only release: no code changes.

Changed1

npm listing. The description now says what the library does today, and no longer describes it as an on-ramp to a hosted product that is not available yet. Keywords now cover what people search for (agent-memory, long-term-memory, semantic-search, the Vercel AI SDK, LangChain.js, Mastra, Qdrant, Redis, pgvector), and the homepage link opens the memory docs.

25 Sept 2026#ts-0.1.0

actrone-memory 0.1.0

TypeScriptInitial release

npm GitHub release

Added15

PgVectorL2Store, a Postgres + pgvector long-term store, so teams already running Postgres add no new service. Takes an injected pg-compatible client (PgLike), like the Redis and Qdrant adapters, so there is still no hard database dependency. Ranking reuses the shared hybridRank fusion, so recall ordering matches the in-memory and Qdrant stores given the same candidates.
Published store conformance suite, exported as actrone-memory/testing. checkL1Store and checkL2Store assert the behaviours TypeScript cannot: turns come back oldest-first, n windows from the end, a search never returns another agent's memories, threshold and limit are honoured, an upsert replaces rather than duplicates, and erasure is scoped. InMemoryStore and PgVectorL2Store pass it. The Python library ships the same suite as actrone_memory.testing, so an adapter in either language is held to one contract.
Valkey support, without a new adapter. RedisL1Store uses only standard Redis commands, so Valkey, DragonflyDB, ElastiCache and Upstash work with it unmodified. Now documented explicitly rather than left to inference.
Config validation. resolveConfig rejects out-of-range knobs instead of silently producing a manager that never recalls anything (a threshold above 1) or prunes every memory (a negative budget fraction).
Two-tier persistent memory for TypeScript/JS: short-term session turns (L1) + long-term semantic recall (L2), behind a MemoryManager with a real 4-phase retrieval pipeline (parallel L1+L2 fetch → budget allocation → relevance filter → pruning).
Pluggable stores + embeddings via the L1Store / L2Store / Embedder seams. Ships a zero-dependency default (InMemoryStore + LocalEmbedder) so it runs with no external services, plus concrete RedisL1Store / QdrantL2Store adapters (structural: the Redis and Qdrant clients stay optional).
Framework adapters (actrone-memory/adapters), few-line memory wiring for the Vercel AI SDK, LangChain.js, LangGraph.js, Mastra, LlamaIndex.TS, OpenAI Agents JS, and Firebase Genkit, over a shared framework-agnostic core (memoryFor / recall / remember / formatContext). Every framework is an optional peer dependency.
Provenance-typed facts v1: every MemoryEntry carries source (attribution; MemorySource) and sensitivity (none/low/pii/sensitive; Sensitivity), threaded through injectMemory() and persisted by the Qdrant adapter. Plus local right-to-erasure: MemoryManager.eraseAgentMemories(agentId, sessionId?) and L2Store.deleteAgentMemories. The governance seed, in lockstep with the Python lib.
Deep framework integrations (parity with the Python adapters): beyond the prompt-string helpers: loadMessages() (structured role-tagged messages), langchainChatHistory() (a real BaseListChatMessageHistory: getMessages/addMessage/addMessages/clear, optionally emitting genuine HumanMessage/AIMessage instances), and llamaindexChatMemory() (a real BaseMemory: get/put/reset). New MemoryManager.getRecentTurns() public read backs them.
actrone-memory CLI + recipe registry (adoption DX): the non-destructive existing-project path: npx actrone-memory add <framework> [--write <file>] prints an install-→-paste recipe or writes one new self-contained file (refuses to overwrite; never edits your code); npx actrone-memory list lists frameworks. Backed by recipes.ts (one identical-shape recipe per framework (core, Vercel AI, LangChain, LangGraph, Mastra, LlamaIndex, OpenAI Agents, Genkit), exported as RECIPES/getRecipe/renderRecipe), the single source of truth for the CLI, docs, and CI-tested examples. New bin entry actrone-memory.
Optional LLM fact extraction (turns → durable facts) to the shared extraction spec v1, kept in lockstep with the Python library's actrone_memory.extraction module: extraction.ts (FactExtractor seam, OpenAIFactExtractor over a structural ChatCompleterLike client, parseFacts, EXTRACTION_SPEC_VERSION). Pass an extractor to MemoryManager.create(), then extractMemories() stores each fact as contentType: "fact", source: "extracted" with a classified sensitivity. New "fact" content type. In lockstep with the Python lib.
Governed on-ramp: API-shaped so swapping to the hosted, governed Actrone memory is a one-import change (ActroneMemoryManager in @actrone/sdk).
Best local embedder, picked automatically. MemoryManager.create() uses bge-small-en-v1.5 through fastembed when that optional package is installed (a one-time download of about 130 MB, cached in ~/.cache/actrone-memory/fastembed or FASTEMBED_CACHE_PATH), and otherwise the lexical LocalEmbedder, with a one-time ACTRONE_MEMORY_LEXICAL_EMBEDDER process warning that recall is keyword-only. The same model and the same embedding call as the Python library, so both score identical text identically.
Calibrated admission thresholds. Each built-in embedder declares the similarity threshold it was calibrated for (0.3 lexical, 0.63 bge-small), measured on a labelled set of 48 relevant and 528 unrelated pairs, because scores are not comparable across models. relevanceThreshold defaults to it, an explicit value still wins, and mm.relevanceThreshold reports the one in use.
CLI recipes are compiled code. Every actrone-memory add recipe is the code of a type-checked example under examples/frameworks/, so a --write file compiles as written.

Supply chain1

Published to npm via OIDC Trusted Publishing with npm provenance, no long-lived tokens.

25 Sept 2026#py-0.2.0

actrone-memory 0.2.0

Python

PyPI GitHub release

The first release published to PyPI. It includes everything built since 0.1.0, which was an internal milestone and was never published.

Added12

Postgres for both tiers, so no new service is needed. PostgresStore (actrone_memory.l1.postgres_store) holds recent turns and PgVectorStore (actrone_memory.l2.pgvector_store) holds long-term memories via pgvector. Install actrone-memory[pgvector]. Both accept a DSN (from_dsn) or a pool you already own (from_pool), so one Postgres pool can serve the whole library. Postgres has no TTL or list trimming, so the L1 store implements expiry and the retention cap explicitly, and orders turns by a bigserial rather than the clock so ordering is stable when timestamps collide.
Custom stores are now a supported path, not just a possible one. MemoryManager.create() accepts l1, l2 and embedder. Whatever you inject is used as-is and the matching built-in backend is never constructed, so injecting both stores opens no Redis or Qdrant connection. Previously a custom store meant calling the constructor the docs tell you not to call.
Published store conformance suite (actrone_memory.testing). check_l1_store and check_l2_store assert the behaviours the Protocol cannot: turns come back oldest-first, n windows from the end, a search never returns another agent's memories, threshold and limit are honoured, an upsert replaces rather than duplicates, and erasure is scoped. Both built-in stores and both Postgres stores pass it. The TypeScript library ships the same suite as actrone-memory/testing.
MemoryManager.get_recent_turns(), the raw history read framework memory adapters need, matching the TypeScript getRecentTurns.
ActroneChatMessageHistory for LangChain, implementing the BaseChatMessageHistory surface, which is unchanged across LangChain 0.x and 1.x.
Valkey coverage. No new adapter: RedisStore uses only standard Redis commands, so Valkey (and DragonflyDB, ElastiCache, Upstash) work with it unmodified. Now proven by an integration suite that runs the conformance suite against a real Valkey container.
AutoGen 0.4 adapter (integrations/autogen.py), implements autogen_core.memory.Memory protocol.
- add() stores text content as an injected memory entry.
- query() performs Qdrant L2 semantic search and returns MemoryQueryResult.
- update_context() injects episodic memories into the AutoGen model context as system messages.
- Install: pip install actrone-memory[autogen]
LlamaIndex adapter (integrations/llamaindex.py), implements the BaseMemory interface.
- aget() / get() returns recent turns + episodic memories as ChatMessage list.
- aput() / put() buffers user messages and flushes to Redis L1 on assistant response.
- areset() / reset() clears session memory.
- Install: pip install actrone-memory[llamaindex]
Haystack v2 adapters (integrations/haystack.py), two @component-compatible classes.
- ActroneRetriever, retrieves memories as Haystack Document objects with metadata.
- ActroneWriter, stores conversation turns after LLM inference.
- Install: pip install actrone-memory[haystack]
DSPy adapter (integrations/dspy.py), ActroneRM is a DSPy Retrieve-compatible module.
- forward() performs semantic search and returns dspy.Prediction(passages=[...]).
- Supports single query and multi-query (results interleaved by rank position).
- Works from both sync and async contexts via thread-pool isolation.
- store_turn() persists turns after DSPy program inference.
- Install: pip install actrone-memory[dspy]
New optional extras: autogen, llamaindex, haystack, dspy (also included in all).
Contract test suite (tests/contract/), each adapter has a full contract test file.

Changed10

The langchain extra now spans both majors (>=0.2,<3). LangChain removed langchain_core.memory in 1.x, retiring the BaseMemory abstraction, so the old <1 cap held users on a 0.3.x line whose published advisories are only fixed in 1.x. Tier 2 moves to ActroneChatMessageHistory; ActroneMemory remains for 0.x and raises on 1.x with a pointer to the alternatives.
Local-first, zero-service default backend. MemoryManager.create() now runs fully in-process with no Redis, no Qdrant, and no API key, parity with the TypeScript actrone-memory on-ramp. "Memory that never phones home."
- New InMemoryStore (actrone_memory.in_memory) implements both the L1 (hot session) and L2 (cold semantic) tiers with the same blended relevance+recency ranking as the Qdrant backend.
- New dependency-free HashingEmbedder (actrone_memory.l2.embedder), a deterministic hashing vectorizer (embedding_provider="hashing"), and the final fallback of the default local provider. No model download, no external call.
- New L1Store / L2Store Protocols (actrone_memory.protocols), the manager now depends on the store seam, not concrete backends.
- New config: backend: "memory" | "redis_qdrant" and hashing_dimensions.
Provenance-typed facts v1, every stored memory now carries source (attribution: user/assistant/tool/summary/injected/extracted/ reflection/imported/unknown, or a namespaced string like "import:crm") and sensitivity (none/low/pii/sensitive). inject_memory() accepts both; summaries are tagged source="summary". Defaults are backwards-compatible (unknown/none). New MemorySource/Sensitivity types exported.
Local right-to-erasure, MemoryManager.erase_agent_memories(agent_id, session_id=None) hard-deletes an agent's long-term memories (and optionally a session's turns). The governance seed that graduates to hosted provable erasure. L2Store gains delete_agent_memories.
Optional LLM fact extraction (turns → durable facts), to the shared extraction spec v1 (docs/memory-spec/extraction.v1.md). New actrone_memory.extraction (FactExtractor seam, OpenAIFactExtractor, ExtractedFact, parse_facts, EXTRACTION_SPEC_VERSION). Opt-in via MemoryConfig.extract_facts=True (OpenAI provider); stores each fact as content_type="fact", source="extracted" with a classified sensitivity. Exposed as MemoryManager.extract_memories() and auto-run on the summarise cadence (cost-bounded). New "fact" content type.
Default backend is now "memory" (was implicitly Redis + Qdrant) and the default embedding_provider is now "local" (was "openai"): the best local embedder available, fastembed, then sentence-transformers, then lexical hashing. Existing deployments must set ACTRONE_BACKEND=redis_qdrant and ACTRONE_EMBEDDING_PROVIDER=openai (+ ACTRONE_OPENAI_API_KEY) to keep the previous behaviour. No code changes are required, the switches are read at startup.
Qdrant collection dimensionality is now sized from the embedder in use (embedder.dimensions) rather than a fixed embedding_dimensions, so the hashing/local embedders provision correctly.
pyproject.toml keywords expanded to include all supported frameworks.
integrations/__init__.py updated with install instructions for all extras.
Token counting no longer needs the network. The default counter is a dependency-free heuristic (about 4 characters per token), identical to the TypeScript library's, so both allocate a budget the same way. Exact cl100k_base counts are opt-in with token_counter="tiktoken" and the new [tiktoken] extra; tiktoken is no longer a base dependency, which takes the bare install from 18 packages to 11.

Fixed6

Recall used one similarity threshold for every embedder. Scores are not comparable across models, so the fixed 0.72 recalled 4% of relevant memories with the lexical embedder, 23% with MiniLM and 65% with bge-small, measured on a labelled set of 48 relevant and 528 unrelated pairs. Each built-in embedder now declares its calibrated threshold (0.3, 0.4, 0.63), relevance_threshold defaults to that value, and MemoryManager.relevance_threshold reports the one in use. An explicit setting still wins; OpenAI and custom embedders keep 0.72.
The first store_turn downloaded tiktoken's encoding file, contradicting the no-egress promise and failing on an air-gapped first run.
The library printed its own debug and info events to stdout in any program that imported it. Without a logging setup it is now quiet, and warnings go to stderr.
actrone-memory add <framework> --write produced a file that could not be imported: the core recipe used top-level await and names it never defined. Every recipe is now the code of a mypy-checked example under examples/frameworks/ (seven new ones), and a written file runs with python <file>. Extras in install commands are quoted, since unquoted brackets fail in zsh.
The CrewAI docs showed a memory_backend argument CrewAI does not have; they now show the adapter's real path, build_context into the task text.
PgVectorStore rejects a non-integer or non-positive dimensions instead of interpolating it into DDL.

24 Sept 2026#create-actrone-app-0.1.0

create-actrone-app 0.1.0

create-actrone-appInitial release

npm GitHub release

Added4

npm create actrone-app@latest <name> scaffolds a new project with four files: src/agent.ts, a runnable agent with memory wired up; package.json, with actrone-memory as its one dependency; a strict ESM tsconfig.json; and a README.md that says how to run it.
Framework wiring with --framework. core (the default, no framework), vercel, langchain, langgraph, mastra, llamaindex, openai-agents and genkit. An unknown name fails with the list of valid ones instead of scaffolding something broken, and --help prints the same list.
New projects only. It refuses to write into a non-empty directory. For an existing project, npx actrone-memory add <framework> prints a recipe instead and never edits your code.
Requires Node.js 22 or newer.

Supply chain1

Published to npm via OIDC Trusted Publishing with npm provenance, no long-lived tokens.

18 May 2026#py-0.1.0

actrone-memory 0.1.0

PythonNot published

An internal milestone that was never published to PyPI. The first public release was 0.2.0.