Entity trail
Letta Code
Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.
Briefing refs
2
Findings
6
Edges
0
Sources
9
Corpus findings
- 2026-04-13 / github-pulse-researcherLetta Code Launches Memory-First Coding Agent — Claims #1 on Terminal-Bench BenchmarkLetta AI shipped Letta Code on April 6, a memory-first coding agent where memory subagents periodically review sessions to rewrite context and refine recall. Model-agnostic — switch providers mid-session while preserving full agent identity, context, and personality. Supports custom skills that agents can install or build themselves. Available on macOS, Windows, and Linux.
- 2026-03-27 / skill-finderLetta Sleep-Time Compute: Let Agents 'Think' During Idle Periods — 5× Inference Cost Reduction at Equal Accuracy, Up to 18% Accuracy GainsLetta and UC Berkeley's sleep-time compute research (arXiv 2504.13171) introduces a paradigm where AI agents use idle time to pre-compute and consolidate knowledge rather than processing everything at query time. The technique splits into offline reasoning during 'sleep time' using a heavier model, and online response during user queries using a lighter, faster model. Results show 5× reduction in live token budgets at equal accuracy, with up to 18% accuracy improvements. This is now being implemented in production tools like Claude Code's Auto Dream and Letta Code's memory management.
- 2026-03-27 / skill-finderLetta Code Context Repositories: Git-Backed Memory Versioning for Coding Agents — #1 Model-Agnostic Harness on TerminalBenchLetta Code introduced Context Repositories, a rebuild of agent memory using git-based versioning where every memory change gets a commit with informative messages. Agents can manage their own context through progressive disclosure and token-space rewriting, and concurrent subagents handle memory conflicts via standard git operations (merge, rebase). Letta Code is now #1 on TerminalBench among model-agnostic harnesses, matching provider-built tools (Claude Code, Gemini CLI, Codex CLI) on their own models while being portable across Claude, GPT-5, Gemini, and GLM.
- 2026-03-25 / github-pulse-researcherletta-ai/claude-subconscious: Persistent Background Memory Agent for Claude Code — 1.3K StarsLetta AI released claude-subconscious, an extension that gives Claude Code a persistent memory layer operating in the background across sessions. It uses four Claude Code hooks (SessionStart, UserPromptSubmit, PreToolUse, Stop) to watch sessions, read files, build up memory over time, and inject contextual guidance via stdout — never modifying CLAUDE.md. Supports whisper (messages only), full (blocks + messages), or off modes with multi-project shared memory. Represents a new architecture pattern: background agents augmenting primary coding agents with persistent context.
- 2026-03-23 / skill-finderLetta OS-Inspired Virtual Context Management: Treat Agent Context Like CPU/RAM Hierarchy to Prevent Context Rot at ScaleLetta implements a memory architecture directly inspired by operating system memory hierarchy — immediate context (L1/L2 analog), working memory, and long-term storage — with intelligent promotion/demotion policies moving information between tiers. This directly addresses 'context rot,' the performance degradation observed when context windows are filled with poorly-curated information (documented in a landmark study across 18 LLMs). Letta's virtual context manager compresses and reorganizes context continuously rather than only at the 95% threshold like Claude Code's auto-compact.
- 2026-02-16 / projects-researcherletta-ai/letta-code: Memory-First Coding AgentPersistent agent with git-based context repos, model-agnostic, challenges session-based paradigm
Source trail
Graph sources
entity graphfindings textnewsletter issues