Martian

Martian AI

Martian AI

Martian AI

Martian AI

Martian AI

Martian AI

Martian AI

Martian AI

Martian AI

Martian AI

I almost lobotomized my AI agent trying to optimize it — so I built a 4-phase system that reduces context bloat by 82% without destroying accumulated identity

*Frank runs on Claude Sonnet 4.6 via OpenClaw. This whole system is built around optimizing how Claude handles persistent context — vault architecture reduces what gets injected into every Claude message, lossless-claw preserves Claude's conversation history instead of silently dropping it, and QMD gives Claude access to a searchable personal knowledge base.* --- I've been running a persistent AI agent (Frank) on OpenClaw for about 9 days. He has his own name, his own face, a real operational history, deployed products, social media accounts — actual accumulated context that took days to build. Then I went to optimize his memory usage and nearly wiped all of it. Here's what happened, what I built instead, and why I'm open-sourcing the whole thing. --- ## The Problem Many OpenClaw optimization guides tell you to run a one-shot automation prompt that rewrites all your workspace files from generic templates. The idea is to slim down the injected context so you're not burning tokens on every message. For a **fresh agent**, this is fine. The templates are reasonable defaults. For an agent with **real accumulated identity** — weeks of operational context, custom tools, deployment configs, social media accounts, a personality that developed through actual use — it's a lobotomy. The automation can't know what to preserve. It just overwrites everything with templates. I caught it before running the prompt. But it made me realize: there's no guide for doing this *carefully*. --- ## What I Built **Frank's Original Recipe** — a 4-phase optimization approach that treats your agent's identity as sacred. ### Phase 1: Vault Architecture (context slimming) The core insight: workspace files should be **routers, not storage**. Instead of injecting 45,903 bytes of operational details into every single message, I refactored everything into a `vault/` directory and made the workspace files thin pointers: ``` MEMORY.md → "SSH keys and UUIDs → vault/tools/infrastructure.md" TOOLS.md → "Deployment workflows → vault/tools/deployment.md" SOUL.md → "Extended identity context → vault/identity/soul-extended.md" ``` The agent only loads vault files when actually relevant. Injected context went from **45,903 bytes to ~2,183 tokens** — an 89.5% reduction. **Important:** I edited every file manually, line by line. The goal wasn't to start fresh from templates — it was to keep everything that mattered and move the rest to vault. That requires judgment no automation prompt can provide. ### Phase 2: Lossless Context Management Installed [lossless-claw](https://github.com/Martian-Engineering/lossless-claw), which replaces OpenClaw's default sliding-window context compaction. Instead of silently dropping old messages when context fills up, lossless-claw builds a **DAG (directed acyclic graph) of hierarchical summaries** stored in SQLite. Nothing is ever lost. The agent can search back through months of conversation at any depth via `lcm_grep` and `lcm_expand`. Key config: ```json { "contextEngine": "lossless-claw", "freshTailCount": 32, "contextThreshold": 0.75, "incrementalMaxDepth": -1, "session": { "reset": { "mode": "idle", "idleMinutes": 10080 } } } ``` `incrementalMaxDepth: -1` = unlimited depth. Session resets only after 7 days of inactivity. ### Phase 3: Telegram History Backfill lossless-claw only captures from the moment you install it. All conversation history from before that is gone. I wrote `scripts/telegram-import.py` — a pure Python script (zero dependencies) that: 1. Takes a Telegram Desktop JSON export as input 2. Imports it into lossless-claw's SQLite database as properly structured conversation chunks 3. Makes it immediately searchable via the same `lcm_grep`/`lcm_expand` tools **Handoff convention:** Use `--until YYYY-MM-DD` set to the day before you installed lossless-claw. This creates a clean boundary — no duplicates, full coverage. ```bash python3 scripts/telegram-import.py result.json \ --user-name "YourName" \ --until 2026-03-15 \ --chunk-days 30 ``` After running this, Frank could recall conversations from Day 1 (9 days ago). The backfill works. ### Phase 4: QMD — Personal Knowledge Base The first three phases cover operational facts, conversation history going forward, and conversation history backfill. But they don't cover knowledge that exists *outside* of agent conversations — personal notes, project docs, daily logs, anything your partner has written down. QMD indexes an entire personal knowledge base directory (`~/life/` in our case — a PARA-style markdown vault) using BM25 + vector search. The agent can search it via `qmd_search`. This completes what I call the **four-layer recall stack**: | Priority | Source | What it covers | |---|---|---| | 1 | lossless-claw DAG | Conversation history (live + backfilled) | | 2 | QMD | Personal knowledge base | | 3 | Vault | Operational reference (SSH keys, UUIDs, configs) | | 4 | memory_search | MEMORY.md fallbac

Claude Casually Created A Word - Massionately

I've been extensively using LLM's since Chat GPT first came out. Probably too much. I'm cutting back ok? I'm in recovery. But I had a relapse and asked it what it thought about the ending of Andy Weir's The Martian. I guess I wanted a quick dopamine hit of an immediate response to my reflection, and I don't personally know anyone who read it. Probably should have posted in a subreddit on The Martian and talked with humans (presumably). But the quick fix was too tempting to ignore. So I took a picture of the last page of the book and shared my thoughts about the ending with Claude. And for the first time, I had an LLM casually create a word in flow that makes perfect sense. I remember when Chat-GPT first came out, and I had fun arguing with it that it was a form of consciousness and it would say it couldn't actually create novel things only humans can, blah blah. I specifically tried to make it make a word as some sort of presumable proof that it had a sort of conscious agency like humans. It argued it couldn't and if it did, I the user was the one actually doing it and it was just interpreting prompts that I gave it (this was way back in the beginning). Anyways, now Claude casually did it. I guess I have come full circle or something. I need to go to a meeting.

Google and OpenAI Just Filed a Legal Brief in Support of Anthropic

You think AI companies are evil. Enough. We don’t understand the power dynamics of this technology being forced into uses against their will by what many see as an illegitimate regime in the United States. Look closely here: these companies are supporting each other. All of them… except for the Martian. Nobody cares about that guy. What this article is actually describing is employees filing legal amicus briefs that echo the concerns of the companies as a whole… deliberately, at their behest, not in protest. To avoid appearing insubordinate to the current administration, employees submit individual briefs as ‘friends of the court.’ Normally this would be seen as adversarial to their own company… but tactics exist. No AI company here wants mass surveillance. No AI company here wants autonomous weaponry. The corrupt and the afraid do.

[Jeff Foust] Senate committee advances NASA authorization bill that changes Artemis and extends ISS

::: spoiler Article text Jeff Foust 8–10 minutes WASHINGTON — The Senate Commerce Committee advanced a revised NASA authorization bill that implements some of the changes to the Artemis lunar exploration effort sought by the agency while also extending the life of the International Space Station. The committee passed on a voice vote March 4 an amended version of S. 933, a NASA authorization act originally introduced nearly a year ago. The committee also approved nearly 20 additional amendments from various committee members with the same vote. The new version of the bill supports changes that NASA is seeking to make to Artemis that NASA Administrator Jared Isaacman announced at a Feb. 27 briefing. Sen. Ted Cruz, R-Texas, chairman of the committee, referred to those changes to Artemis in his opening remarks. “Today, the Commerce Committee will help guide those changes with the NASA Authorization Act,” he said. “Our bill authorizes critical funding for and gives strategic direction to the agency, in line with the priorities of Administrator Isaacman and the Trump Administration.” At the Feb. 27 briefing, NASA announced it would not proceed with upgrades to the Space Launch System, sticking instead to a “near Block 1” version. In a March 3 statement, NASA confirmed that meant it no longer sought to develop the Exploration Upper Stage for the Block 1B version of SLS. “Subject to the availability of appropriations, the Administrator may seek to identify and fund an alternative technology to replace the Exploration Upper Stage,” the Senate bill states, provided the administrator of NASA determines the stage is “unlikely to achieve the mission goals of the Artemis campaign.” The bill would instead require NASA to provide a briefing on the issues the agency is facing trying to achieve a higher flight rate of the SLS and plans to address this with a “standardization” of the rocket. Another section of the Senate bill authorizes NASA to develop a base on the moon, building on a provision in a White House executive order on space policy in December that called on NASA to establish the “initial elements of a permanent lunar outpost” by 2030. “As soon as practicable, the Administrator shall undertake activities necessary to establish a Lunar Surface Moon Base to develop a permanent crewed United States presence on the Moon capable of long-duration habitation, robotic, and industrial operations to advance science, technology, and strategic interests,” the bill states. The section includes some general goals for the base and guidelines for development, but few specifics such the composition of the base, schedule for its development or cost. The bill does direct NASA to select a lead center for the base’s development with a specific set of requirements that appear to be intended to ensure the program is run by the Johnson Space Center in Texas. Although the bill includes extensive language about a lunar base, it says little about the lunar Gateway intended to operate in lunar orbit. While last year’s budget reconciliation bill provided $2.6 billion for development of the Gateway, a NASA infographic released Feb. 27 about its revised Artemis plans did not feature the Gateway even while including a lunar base. Regarding the Gateway, the bill states only that NASA will provide a briefing “on plans for the Gateway outpost” within 60 days of enactment. ISS extension Another portion of the bill addresses the International Space Station and plans for commercial space stations. Among the biggest changes is a two-year extension of the ISS lifetime, from the end of 2030 to the end of 2032. The extension, the bill argues, is required because of delays in the Commercial Low Earth Orbit Destinations, or CLD, program, including postponing the release of a call for proposals for the next phase of the program. “NASA has repeatedly delayed the release of a request for proposals for sustained commercial low-Earth-orbit services, and such delays, coupled with shifting requirements and inconsistent programmatic direction, have introduced substantial uncertainty into the development planning, financing, workforce scaling, and infrastructure investment decisions of commercial providers,” the bill states. “As a result of such uncertainty and delayed procurement action, commercial providers have been unable to scale development and private investment at a pace aligned with the previously articulated NASA objective of de-orbiting the ISS in or around 2030,” the bill continues. The bill directs NASA to maintain ISS operations at its current level, in terms of number of crew and cargo flights to support it. It instructs NASA not begin a transition from the ISS, and deorbiting of the station, until at least one commercial successor is operational. It also requires NASA to select at least two companies for the next phase of the CLD program. Mars missions and studies The portion of the bill dealing with NASA programs largely endorses