TypeScript agents that remember, forget, and write their own tools

What it does

AgentOS is a TypeScript framework for building AI agents that maintain persistent cognitive memory across sessions, optionally express HEXACO personality traits that bias retrieval and routing decisions, and expand their own capabilities by generating TypeScript tools at runtime. It provides a unified dispatch interface across eleven LLM providers and includes multi-agent orchestration, streaming guardrails, and a voice pipeline. The memory layer implements eight neuroscience-backed mechanisms—including Ebbinghaus decay and retrieval-induced forgetting—so agents forget information in a roughly human-like way rather than retaining everything or dropping it randomly.

The interesting bit

Instead of hallucinating a workaround when its toolkit falls short, an agent generates a TypeScript function with a Zod schema, submits it to an LLM judge for approval, and runs the result in a hardened node:vm sandbox before adding it to a session-wide index. The framework also treats markdown as the source of truth for long-term memory via SOUL.md workspace files and a memory/ wiki that rebuilds vector and graph indexes from plain text.

Key highlights

• Benchmarked at 85.6% on LongMemEval-S and 70.2% on LongMemEval-M (1.5M-token haystacks), with published methodology.
• Runtime tool generation with LLM-judged approval and hardened node:vm sandboxing (5-second timeout, 128 MB heap limit, banned eval/require by default).
• Optional HEXACO personality vectors that live in the kernel rather than the prompt, biasing retrieval and specialist routing across sessions.
• SOUL.md workspace files plus a markdown memory/ wiki for human-readable, versionable long-term memory.
• One dispatch interface across 11 LLM providers (9 API-key services plus 2 local CLI options, with OpenRouter access to 200+ models).

Caveats

• The default node:vm sandbox is explicitly flagged as a tradeoff; the docs note that multi-tenant hosting requires the stronger isolated-vm queue instead.
• HEXACO personality is opt-in, and the README admits most production deployments do not pass a trait vector—suggesting the feature is more niche than the feature list implies.

Verdict

Worth evaluating if you are building long-running TypeScript agents where memory fidelity and dynamic tool growth matter more than a static playbook. Skip it if you just need a thin LLM wrapper—this is a heavy framework with strong opinions about memory, identity, and sandboxed code generation.