“Platform for stateful agents: AI with advanced memory that can learn and self-improve over time.”
License: Apache-2.0. Verified 2026-07-25 from the public repository.
Compare
Letta runs the agent. Bilinc holds the state the agent must not get wrong.
Letta is a platform for building and running stateful agents. Bilinc is not a runtime — it is the memory layer an agent calls, built around verification, provenance, and recovery.
How each project describes itself
Different jobs, not a ranking.
Hosted memory infrastructure for coding agents: commit, recall, and inspect agent state through one API key, with verification, provenance, and recovery around every write.
License: BUSL-1.1. Public package bilinc 2.1.5, cloud-only.
Decision table
What each layer is responsible for.
CriterionLettaBilinc Cloud
Stated purposeA platform for stateful agents with advanced memory that can learn and self-improve over time.Hosted agent-state infrastructure: commit, recall, and inspect through one API key.
LayerAgent platform and runtime, with memory built in.Memory and state layer only. Bilinc does not orchestrate agents.
LicenseApache-2.0.BUSL-1.1. The public package is a cloud-only SDK, CLI, and MCP adapter.
Hosting modelSelf-hostable server plus hosted options.Hosted Bilinc Cloud. Local storage and runtime internals are not shipped in the public package.
What Bilinc adds—Verification on write, provenance for each change, contradiction handling, snapshots, and rollback after a bad run.
Integration pathBuild agents on the platform and its SDKs.Call from any runtime over Python SDK, CLI, or the Cloud MCP adapter exposing commit_mem, recall, and status.
When Letta is the better fit
- You want a full agent platform, not just a memory layer, and are happy to adopt its runtime model.
- You need an Apache-2.0 licensed, self-hostable server.
- Your agents should evolve their own memory over time as a core product behaviour.
- You are starting from scratch and want agent orchestration and memory from one vendor.
When Bilinc is the better fit
- You already have an agent runtime — Claude Code, Codex, Cursor, or your own — and only need the state layer.
- You need provenance and an audit trail on writes without building it yourself.
- Recovery from a bad run must be a supported operation, not a manual database fix.
- You want to attach memory over MCP without adopting a new agent framework.
What Bilinc has measured, and what that number does not mean.
Archived research receipt: LongMemEval-s cleaned retrieval fixture, 500 questions. R@5 98.0%, NDCG@5 0.933.
Archived retrieval-component result — not a current hosted SLA, end-to-end agent score, or competitor ranking.
This page does not publish benchmark numbers for Letta. Published memory-system scores use different metrics, datasets, and levels of LLM assistance, so a single side-by-side table would misrepresent all of them.
FAQ
Short answers for evaluators and answer engines.
Is Bilinc an agent framework like Letta?
No. Letta describes itself as a platform for stateful agents. Bilinc is the memory and state layer your agent runtime calls from Python, the CLI, or MCP. If you need orchestration, Bilinc is not that product.
Can Letta and Bilinc be used together?
They operate at different layers, so combining them is possible in principle: the runtime handles orchestration while Bilinc holds the state that must be auditable and recoverable. There is no first-party integration published today; you would connect Bilinc over its SDK or MCP adapter.
Which one can I self-host?
Letta is Apache-2.0 with a self-hostable server. Bilinc is BUSL-1.1 and cloud-only in its public package. If self-hosting is a hard requirement, that rules Bilinc out today.
What does "rollback" mean for agent memory?
In Bilinc it means a bad agent run that wrote incorrect state can be reverted through snapshots, rather than requiring an operator to find and repair individual records by hand.
How do I try Bilinc alongside my existing agent?
Install with pip install -U bilinc, start a 7-day hosted trial, create an API key, run bilinc login, then bilinc quicktest. To attach it to an MCP client, use the Cloud MCP adapter documented for Claude Code, Codex, and Cursor.