engram: A Recursive Language Model Engine for Coding Agents

Manav Arya Singh · Independent · May 2026

Abstract

Long-context coding agents fail in two specific ways. First, frontier models with one-million-token windows lose 25 to 60 percentage points of retrieval accuracy past 200K tokens, an effect documented in 2026 needle-in-haystack benchmarks. Second, multi-turn sessions accumulate tool outputs and intermediate reasoning until compaction discards most of what the agent learned. Existing agent-memory systems — Mem0, Letta, Zep — address the second by storing chat memory, but none expose primitives appropriate to source code, and none are verifiable in the sense that the agent's traversal can be reconstructed from a cryptographic record.

In December 2025, Zhang, Kraska, and Khattab (MIT CSAIL) introduced Recursive Language Models, a paradigm in which long context lives in an external environment that the model examines through a REPL. The reference implementation is general-purpose. We present engram, the first specialization of that paradigm for coding agents. engram exposes five primitives — grep, read, ast, git, recurse — chosen to match the queries a coding agent actually issues. Every primitive funnels through a single byte-budgeted runner that appends one JSON Lines entry per call to a local journal: timestamp, primitive, arguments, SHA-256 hash over key-sorted JSON output, bounded preview, duration. The journal is append-only and local-only by construction, providing a verifiable record of what the agent examined at any past moment.

engram ships as a Claude Code plugin and a standalone CLI. The codebase is never loaded into the model's context window; the model navigates programmatically through the primitive REPL. v0.0.1 implements two of the eight memory tiers identified in the architecture — the journal (L2) and the code-as-environment view (L4) — with the remaining six on the roadmap. The full source, design contracts, and threat model are in the project repository.

Citation

@software{singh2026engram,
  author = {Singh, Manav Arya},
  title  = {engram: A Recursive Language Model Engine for Coding Agents},
  year   = {2026},
  url    = {https://github.com/Manavarya09/engram}
}

The Recursive Language Model paper this work specializes:

@article{zhang2025recursive,
  author  = {Zhang, Alex L. and Kraska, Tim and Khattab, Omar},
  title   = {Recursive Language Models},
  journal = {arXiv preprint arXiv:2512.24601},
  year    = {2025}
}

Source markdown: docs/paper/engram.md · by Manav Arya Singh.