The State of Agent Memory in 2026

The number that started everything

Mem0's npm package — mem0ai — was getting 3,000+ weekly downloads before it even had a working JavaScript SDK. The README pointed to a Python library. People were installing an empty package just because the word "memory" was in the name.

That one metric convinced us to look deeper. If there's this much demand for agent memory, what does the supply actually look like?

By the numbers

We audited 10 open-source agent memory projects spanning three distinct paradigms: vector-based retrieval, knowledge graphs, and hybrid approaches. Collectively, they represent 120,000+ GitHub stars, $31.5M in venture funding, and wildly different bets on how agents should remember.

Key findings

• Graphiti's bi-temporal model is the most architecturally sophisticated approach we found. It tracks both when an event happened and when the system learned about it — essential for correcting stale memories without losing history.

• memU's salience scoring borrows from cognitive science. Memories decay based on recency, frequency, and emotional weight. It's the closest anyone has come to modeling how human memory actually prioritizes information.

• SimpleMem's multi-query retrieval generates multiple search queries from a single user prompt, then merges results. Simple idea, significant accuracy improvement over single-vector search.

• Mem0 fixed a UUID hallucination bug where the LLM was generating fake memory IDs during graph updates. Their solution: strip UUIDs from the prompt entirely and let the system handle references. A reminder that LLMs and structured identifiers don't mix.

• mcp-memory-service implements exponential decay with a configurable half-life. Memories literally fade unless reinforced — the most aggressive temporal model in the landscape.

• Letta (formerly MemGPT) takes the most literal approach: a full operating-system metaphor with main memory, archival storage, and an agent that manages its own context window.

The comparison

The missing layer

Notice the last column. Every project in this landscape is optimized for conversational context — chat history, user preferences, accumulated facts from dialogue. None of them are ingesting real-time engineering signals: commits landing, PRs being reviewed, deployments shipping, incidents resolving.

That's the gap. Your codebase generates a continuous stream of structured events that carry institutional knowledge — why a decision was made, what broke last time, which patterns the team has converged on. No memory framework is capturing it.

This is what we're building with Mneme: a memory layer that starts from real-time signals, not chat transcripts. Architecture decisions extracted from PR reviews. Failure patterns learned from incident timelines. Tribal knowledge surfaced from the code itself.