On July 24th 2026, Letta announced the open‑source trajectory package, a standard format for recording coding‑agent sessions across multiple harnesses such as Claude Code, Codex and Letta Code. The blog post on the Letta research site frames the launch as a response to the fragmented experience data that limits today’s agents from learning across tools.
"Introducing trajectory, an open‑source package that normalizes coding‑agent sessions from Claude Code, Codex, Letta Code, and other harnesses into one token‑efficient format designed for agents learning from past experience," the Letta team wrote in the announcement.
The package defines a trajectory schema – a JSON‑compatible list of records that capture user messages, agent reasoning, tool calls and tool results. The first record is a meta entry that records the originating harness, working directory, git branch and model identifier. Subsequent records follow a simple role‑based structure (user, reasoning, assistant, tool) and include timestamps. By stripping out per‑step envelopes, duplicated payloads and encrypted reasoning blobs, the format reduces the token count of a typical session by roughly 5× compared with native harness logs.
The Letta blog provides a concrete token‑reduction table measured with the Anthropic count‑tokens API on real coding sessions. For a Claude Code session that originally required 951,115 tokens, the trajectory format (default settings) recorded only 170,934 tokens – a 5.6× reduction. Similar gains were observed for Codex sessions, with reductions ranging from 4.5× to 5.6× depending on whether long tool results were truncated.
Developers can use the package via a small TypeScript API. The blog example shows how to list available sessions from any supported source and normalize them into the standard record format:
import { listTrajectories, normalizeTranscript } from "@letta-ai/trajectory";
import { readFileSync } from "fs";
const page = await listTrajectories({ source: "claude-code", limit: 10 });
const transcript = readFileSync(page.items[0].path, "utf8");
const { records, diagnostics } = normalizeTranscript({ source: "claude-code", transcript });
Beyond simple conversion, Letta illustrates how the format powers its dreaming process – a background routine that reviews recent sessions, extracts lessons and writes them into persistent memory. By normalizing sessions from every local harness, the dreaming engine can aggregate experience across Claude Code, Codex and Letta Code, effectively giving a single agent a multi‑tool memory history.
The launch positions Letta as a facilitator of cross‑harness learning at a time when developers increasingly spin up agents from competing providers. Projects such as Harbor’s ATIF have offered full‑fidelity replay formats, but those are geared toward benchmarking rather than efficient ingestion by other agents. Letta’s trajectory format deliberately sacrifices some detail to keep token footprints low, making it more suitable for on‑device memory formation and retrieval.
The package is available on npm as @letta-ai/trajectory
and can be installed globally with npm install -g @letta-ai/trajectory
. Letta also provides a companion CLI for its own code harness – installable via npm install -g @letta-ai/letta-code
– which includes a /init
command to pull in trajectory data from other agents.
The broader AI‑coding ecosystem has long struggled with siloed session logs. By offering an open‑source, token‑efficient canonical format, Letta hopes to catalyze tooling that aggregates experience at scale, ultimately enabling agents to improve from the collective knowledge of many users and platforms.
For the full announcement, see Letta’s blog post Trajectory: A Standard Format for Agent Experience Data.