Leo Lin · 林中玥

Personal memory and personal data for AI agents, plus the infrastructure and hardware they run on.

Local-first memory · human-approved writes before model proposals · on-device inference · k3s home lab

简体中文 · English

One line of work: personal memory and personal data for agents, the perception and judgment on top of it, the agent infrastructure underneath, and the hardware it reaches. The control order is human > mechanism > model.

• A write is applied by a person. A model may propose a change; applying it stays a human step.
• Forgetting is a mechanism, not a model call. Fixed rules choose what is evicted, and each row carries the reason it was chosen.
• An inference is not the user's voice. What a model guessed is capped below the bar that would let it reach a live turn as the user's own statement.
• Negative results stay in the record. Measurements are dated and rerunnable; a result that did not clear its bar is reported as such.

Collection, then structuring and judgment, then the memory layer, then the device. The dates are the periods each stage covered.

flowchart LR
    A["华润万象生活<br/>Collection<br/>2024.01–2025.02"] --> B["小铁<br/>Structure and judgment<br/>2025.03–2026.03"]
    B --> C["BicaMind<br/>Memory and device<br/>2026.05–"]
    C --> D["memfw<br/>Forgetting and retention<br/>2026.10–"]

Personal memory and data, then perception and judgment, then agent infra, then hardware.

• memfw — framework. A model-agnostic, local-first personal memory framework in Python.
• memfw — retention. Five forgetting strategies (fifo, lru, decay, salience, consolidate_then_forget) are pure functions and return a reason per row; human anchors are never evicted.
• memfw — write path. Markdown is the authority, a hash-chained JSONL log records every state change, and SQLite FTS5 is a disposable index. The framework does not call an LLM and does not send data out.
• memfw — local LoCoMo run (2026-10-19). 10 conversations, 5,882 turns, 1,986 questions, 1,000 paired bootstrap samples with Holm correction, finished in 249 seconds. No pairwise difference between strategies is significant; at a 25% budget salience has the highest point estimate, 14.5% against FIFO's 13.0%, and that gap is not significant. On 200 synthetic items, no strategy cleared the collection bar.
• memfw — import pilot. 656 real memories on local MLX (Qwen2.5-7B 4-bit) stalled: after 6+ hours only 498 rows landed, because each save rechecks the whole store. Model proposals are paused, batch writes are next. 38.4% of those records have only a file mtime as a timestamp.
• memfw — checks. 396 tests and mypy --strict passed as of that snapshot.
• BicaMind. Long-term memory for a voice agent. Explicit user memory is stored as such, replayed at confidence 0.95 and recalled next session; inferences are capped at 0.65, under the 0.7 bar of the always-on voice prompt, so they reach a live turn only through prefetch.
• sayless. Local-first dictation, a non-commercial learning project: record → whisper.cpp on device (about 1 second for 8 seconds of speech; audio stays on the machine) → a deterministic dictionary → LLM polish → paste into the focused app. History and the lexicon stay local.
• Day-to-day agent memory. Several agents share one Markdown memory on a private NAS (user / feedback / project / reference / incident; credentials stay out). Snapshot 2026-10-17: about 600 memories, 1,400+ change events, 2,963 commits across several agent identities.
• Companion-model evaluation. 150 scenarios (130 test / 20 dev, 17 subtypes) are frozen, blind-judged, turned into training data, then checked before and after. Real generations cover 9 providers: 1,168/1,170 outputs, 660 pairwise judgments. Still open: a synthetic set, a single judge, 13 rank reversals, an unfilled human-calibration pack, and a 500-step LoRA run.
• Flickfic. A University of Hong Kong student startup initiated by REMi; I am the main engineer. Reader behaviour (views, dwell, drop-off) drives regeneration of the weak states. Not linked here.

• On-device models. MLX, GGUF, and whisper.cpp.
• Home lab, on Tailscale. Multi-GPU machines, Mac MLX for local inference, k3s on three Lenovo Tiny machines, an Apple Silicon node and a NAS. A model farm swaps models as quota and quality change; task state stays in local SQLite.
• Dev boards. RK3566, RK3399, RK3568. Raspberry Pi Pico 2 W (ARM and RISC-V). ESP32.

Star counts below are as stated on 2026-10-20.

• grimoire-skill · ⭐ 8 — parse a PDF once, forge two artifacts: typed reading notes for Obsidian and a reusable skill pack.
• oral-interview-transcription-skill · ⭐ 16 — oral-history recordings → faithful edited transcripts, with an adversarial fidelity gate; data stays local.
• r-analytics-skill · ⭐ 5 — 93 package-level R skills for Claude Code.
• skill-librarian · ⭐ 1 — lint and index a large local SKILL.md library: duplicates, dead links, name collisions, weak descriptions.
• paper-reading-workflow-zh · ⭐ 1 — evidence-grounded Chinese paper-reading workflow (PDF → MinerU → Obsidian).
• cn-cc · ⭐ 1 · mmteam-cc — route Claude Code work to Chinese model backends; multi-model agent teams with fan-out consensus.
• ScreenBar — a SwiftUI menu-bar app that turns Macs on your tailnet into second screens.

Merged pull requests in other people's repositories, coursework set aside. Counted 2026-10-20: 70 merged across the 9 repositories below, 8 still open. Star counts are exact.

In Peter Steinberger's repositories alone: 30 merged PRs (CodexBar 24, oracle 4, RepoBar 2), counted 2026-10-22 — a later snapshot than the table, which shows CodexBar at 22.

Per-PR highlights and the counting rules: docs/upstream-contributions.md.

Cards below are generated by third-party services from the public LeoLin990405 account; the numbers move on their own and are not maintained in this repository.

• 2026-10 — sayless is public; memfw's full LoCoMo run is recorded above, negative result included.
• 2026-09 — agos opened: execution, approvals, and fleet state (#1); honest session, chat, and console slices (#2).

The two workflow files reference third-party actions and credentials that belong to their own projects: grll/claude-code-action and Platane/snk with crazy-max/ghaction-github-pages. Their licences govern those actions.

• GitHub: @LeoLin990405
• Blog: leolin990405.github.io/blog
• LinkedIn: Zhongyue Lin

If one of the projects is useful to you, a star is always appreciated.