LangMem

What Is LangMem?

LangMem is an open-source Python SDK from LangChain that adds long-term memory capabilities to AI agents . Released in early 2025, it provides tooling to extract important information from conversations, optimize agent behavior through prompt refinement, and maintain persistent memory across sessions .

The library builds on LangChain's earlier experiments with hosted memory services and LangGraph's persistent storage layer . It ships as a lightweight pip install langmem package with ~1.5K GitHub stars as of June 2026 .

Status (June 2026): LangMem survived the LangChain 1.0 transition. LangChain's legacy in-process memory classes (ConversationBufferMemory and friends) were deprecated in 1.0, but LangMem was not absorbed or retired — the current docs still position it as the long-term memory option alongside LangGraph checkpointers for session state . The repo is active (commits as recent as June 7, 2026, not archived) , and adoption is real: roughly 746K monthly downloads and 5M+ total on PyPI . The catch is release cadence — the latest PyPI release is 0.0.30, published October 27, 2025, meaning no new package release in over seven months despite ongoing repo activity .

How It Works: Three Memory Types

LangMem organizes long-term memory into three types modeled after human cognition :

Semantic Memory stores facts and knowledge — user preferences, domain knowledge, relationship data. It supports both collections (unbounded searchable stores) and profiles (structured single-document state like a user card). The system handles memory consolidation automatically, merging related facts and resolving contradictions .

Episodic Memory captures past experiences as few-shot examples or conversation summaries. This lets agents learn from successful (or failed) interactions and apply those patterns to new situations .

Procedural Memory modifies the agent's own behavior by updating prompt rules and response patterns. Rather than just storing data, the agent can refine how it operates based on feedback — essentially self-improving its instructions .

Two Paths: Hot Path vs Background

LangMem supports two integration patterns :

• Hot path: The agent calls memory tools (create_manage_memory_tool, create_search_memory_tool) during active conversations to store and retrieve memories in real-time
• Background: A separate memory manager processes conversations asynchronously after they complete, extracting and consolidating memories without slowing the main interaction

Both paths use the same underlying create_memory_manager API and can be combined .

Strengths

• Framework-agnostic core — functional API works with any storage backend, not just LangGraph
• Memory consolidation — automatically deduplicates, merges, and resolves contradictory memories rather than infinitely accumulating
• Structured extraction — supports Pydantic schemas for typed memory profiles, not just raw text blobs
• Native LangGraph integration — works with BaseStore, InMemoryStore, and AsyncPostgresStore out of the box
• Namespace scoping — memories can be isolated per user, shared across teams, or global to the agent
• Open source — MIT licensed, fully inspectable, no vendor lock-in on the core SDK

Cautions

• LangGraph gravity — while technically framework-agnostic, the best experience requires LangGraph and its ecosystem
• Perpetually pre-1.0 — sixteen months after launch the package is still at 0.0.30, and the last PyPI release was October 2025; third-party reviewers call it "newer, less battle-tested" than alternatives like Zep or plain PostgresSaver
• LLM-dependent extraction — memory quality depends entirely on the extraction LLM's ability to identify what matters
• No built-in evaluation — no native tools to measure memory precision/recall or detect memory drift over time
• Managed service adds complexity — the free managed tier is still in signup/waitlist phase, blurring open-source vs. commercial boundaries

What Developers Say

Public developer commentary on LangMem is surprisingly thin given its download numbers — no substantial Hacker News or Reddit threads surfaced as of June 2026. The download-to-discussion gap suggests most usage comes bundled inside LangGraph stacks rather than from teams who evaluated it head-to-head.

The clearest third-party assessment comes from a March 2026 db0.ai survey of post-deprecation LangChain memory options, which lists LangMem as a current Python option but flags its maturity:

"LangMem, LangGraph-native with LLM-driven memory extraction (newer, less battle-tested than the other two [LangGraph PostgresSaver and Zep])" — db0.ai, "LangChain memory is deprecated — what to use in 2026"

Competitive Positioning

Bottom Line

Recommended for: Teams already building on LangGraph who want long-term memory without standing up new infrastructure. It's tightly integrated, well-designed, and the three-memory-type framework (semantic, episodic, procedural) is more thoughtful than most competitors . The procedural memory angle — agents that modify their own prompts based on experience — is genuinely differentiated .

Not recommended for: Teams outside the LangChain/LangGraph ecosystem (Mem0 or Zep offer smoother integration paths), or production systems that need a stable, battle-tested API — the package is still 0.0.x with its last release in October 2025 .

Outlook: LangMem cleared the biggest existential question — it survived LangChain 1.0 as the sanctioned long-term memory layer rather than being absorbed or deprecated , and ~746K monthly downloads show it rides LangGraph's distribution . But the seven-month release gap and missing 1.0 milestone leave its long-term independence ambiguous: it could either graduate to a stable release or quietly fold into LangGraph's core APIs. For LangGraph shops it remains the default answer for agent memory ; everyone else should watch the release cadence before committing.