Engineering Assistant on Letta: Self-Updating Memory, Archival Search, and Portable Workflows
Covers building an agent that independently manages blocks of primary memory, accesses a vector archive, and calls deterministic tools; it also examines shared state across multiple agents, streaming output, and exporting the workflow as a portable file.

Purpose of the assistant
This tutorial builds a stateful engineering assistant on Letta. The agent is given the description “Engineering copilot with inspectable long-term memory.” The architecture combines memory, search, tools, multi-agent interaction, and execution traces.
Criterion: the assistant is suited to stateful engineering tasks and verifiable long-term memory.
Core memory blocks and self-updating
Core memory consists of four verifiable blocks. The agent can update them on its own.

The code defines four blocks with different default values and limits.
| Block | Contents | Default value | Limit |
|---|---|---|---|
persona | “I am Mnemo, an engineering copilot. I am terse, concrete, and I never invent facts — if I don't know something I say so and then go look it up.” | — | 3000 |
human | Persistent information about the engineer: name, role, stack, preferences, and working style. | Unknown so far. | 3000 |
project_state | Current project state: milestones, blockers, decisions, and their rationale. | (empty) | 5000 |
lessons | Lessons learned and mistakes not to repeat; format — one line per lesson. | (empty) | 2000 |
Criterion: the blocks separate the agent's role, the engineer's profile, project state, and lessons learned. The limits define the boundaries of each block.
Embedding, tags, and model configuration
The code sets openai/text-embedding-3-small as the agent's embedding. The code assigns the tags ["tutorial", "copilot"]. The code lists the preferred models: openai/gpt-4.1, openai/gpt-4o-mini, anthropic/claude-sonnet-4-20250514. The system selects the first available model in the list. If none are found, the system uses the first model returned.
Criterion: the list order sets the priority. The fallback is the first model returned.
Programmatic access and agent collaboration
Programmatic access to the blocks serves two purposes.
- state synchronization;
- sharing between multiple agents.
Access is provided to the core memory blocks.
Criterion: shared access to the blocks is needed to synchronize state between agents.
Archival memory and semantic search
Archival vector memory is used for semantic search. It is part of the assistant's architecture.
Criterion: archival search is needed for semantic queries over accumulated data.
Server-side risk assessment tool and streaming execution
A custom server-side tool performs deterministic risk assessment. Streaming execution is part of the assistant's architecture.
Criterion: risk assessment requires a server-side tool with deterministic logic. Execution requires streaming.
Server-side history and portable scenarios
Server-side message history supports persistent conversations. .af export provides portability.
Criterion: persistent conversations require server-side history. Scenario portability requires .af export.
Frequently asked questions
Related materials
All materials
Hybrid Image Annotation: Automation Accelerates Computer Vision Model Training
The article describes how pre-labeling with algorithms reduces manual work and improves the cleanliness of training datasets, as well as why selective human verification remains essential. It covers the method’s limitations in challenging shooting conditions and with rare object types.
5 November 2026
GoldiMask Method for Fine-Tuning Diffusion Language Models: Context Selection and Target Weighting
This paper presents GoldiMask, a supervised fine-tuning approach for discrete diffusion language models that selects visible tokens via approximate maximization of a submodular function and assigns weights to remaining targets based on their training utility. Experiments across several models and datasets demonstrate improved average accuracy, gains on reasoning and coding tasks, and fewer parallel decoding iterations on GSM8K and MATH-500 at comparable quality.
30 October 2026
Transformers adds support for running quantized GGUF models on Apple Silicon
The article describes integrating the GGUF format from llama.cpp into Transformers: local inference on Mac, loading models via the library API, and running an OpenAI-compatible server. Using Qwen3.5-4B as an example, it shows how the choice of quantization level affects weight size and memory requirements.
18 October 2026