A growing class of tools gives LLM agents persistent memory through a filesystem metaphor: ls, cat, grep, and markdown-shaped trees instead of bespoke vector SDKs or unbounded chat context. Three prominent approaches—Wiki CLI, Supermemory SMFS, and Letta MemFS—all support the LLM Wiki / Personal Knowledge vision but optimize for different backends and workflows.
Shared idea
Traditional chat UIs are largely stateless. These systems let an agent (and often a human) read and write durable artifacts as files, compounding knowledge across sessions. They pair well with coding agents (Obsidian Integration as viewer, terminal tools as compiler) but differ in where truth lives and how recall works.
Comparison with Wiki CLI, Letta MemFS, and Supermemory SMFS
Feature matrix
| Wiki CLI | Supermemory SMFS | Letta MemFS | |
|---|---|---|---|
| Metaphor | Semantic markdown wiki / codebase | Cloud container as mount or virtual bash | Git repo of memory markdown |
| Authority | Your wiki on disk | Supermemory API | Local Git (+ Letta Cloud for API agents) |
| Structured semantics | YAML-LD, SHACL, JSON Schema, SPARQL | Memory paths + cloud graph | description frontmatter; optional folders |
| Hot context | Agent chooses pages to open | profile.md + semantic grep |
system/ loaded every turn |
| Search | SPARQL, link graph, wiki query |
Semantic grep (literal with flags) |
Tree + on-demand file read |
| Validation | wiki check (SHACL, JSON Schema, layout) |
Supermemory indexing rules | Git + agent discipline |
| Publishing | Static site, RDF export | N/A (runtime memory) | N/A (agent memory) |
| Typical user | Wiki authors, PKM + semantics | Multi-modal, multi-source agents | Letta Code coding agents |
When to use which
Wiki CLI — You want a local-first, explicit, interlinked wiki that doubles as documentation: Declarative Knowledge in frontmatter, Procedural Knowledge in wiki check / wiki render, optional Retrieval Augmented Generation-style synthesis via SPARQL instead of opaque embeddings. Fits Farzapedia-style gardens with machine-checkable structure.
Supermemory SMFS — You need scale, semantic recall, and integrations (documents, images, Slack, Notion, etc.) with minimal agent onboarding: mount agent_memory/ or drop in @supermemory/bash for serverless. Accept cloud authority and background graph maintenance.
Letta MemFS — You build on Letta Code and want Git history, human-editable memory, and a fixed always-on system/ slice for persona and project facts, with everything else progressively disclosed.
Complementary use
These are not always mutually exclusive. A team might maintain a published Wiki CLI wiki for product knowledge while a Letta agent keeps private MemFS notes, or mount SMFS for ingested corpora that feed summaries written into the wiki. The Wiki CLI’s strength is shared, validated, link-stable markdown; SMFS and MemFS optimize per-agent runtime memory.
Related pages
- Vivary — agent workspace stack (tropo typed graph + strato loop)
- LLM Wiki — pattern origins (Andrej Karpathy, Farza Majeed)
- Second Brain — PKM goals
- Style Guide — conventions for this wiki