studies
ai-labs/studies/context-v/
12 files here · 80 in 6 sub-projects
Sub-projects
agent-harnesses14conversational-ui-and-native-shells12memory-layers-for-agents20open-specs-and-standards16sync-and-content-version-control10vector-databases8plans/4
Analyze Harness-like Evidence in a Collaborator Codebase
Step 1 of 3: read a collaborator's repo for every sign of how they work with coding agents, and inventory it with paths and dates.
Compare and Contrast Harness-like Artifacts between Lossless and a Collaborator
Step 2 of 3: set our harness beside theirs. What's shared, similar, different, missing, and what each side should do about it.
Generate an HTML Page for Lossless and a Collaborator, Rendered in the AI Labs Splash
Step 3 of 3: turn the evidence and the comparison into one splash page. Visual highlights up top, the detailed narrative below.
studies/plans
prompts/3
explorations/4
Create Study Everything as Code
Study Kickoff — Content Authoring Tools
Explore Collaborator Codebases for Insights — Comparative Highlights
When a friend has built their own agentic-coding silo next to ours, compare the silos both ways: what we can learn from them, and what they can learn from us.
studies/explorations
profiles/69
Agent-Native — Profile
The chat isn't bolted onto the app. The app's operations are one action registry that the UI, the agent, MCP, A2A and the CLI all call, and the conversation is just one more caller.
Automerge — Profile
The only entry where two people editing the same paragraph is a solved case rather than a second file with a longer name — and it costs you the filesystem.
git-annex — Profile
Content is optional. git tracks which copies exist where, and refuses to drop the last one — the bookkeeping that makes deleting a local file safe.
Jujutsu (`jj`) — Profile
The storage layer is five pluggable backends, not one — and only two of the five have a shipped implementation you could put in a bucket.
Kopia — Profile
Its snapshot manifest has the free-text Description field restic lacks, and its policies attach per directory and inherit — the ask's own shape, in the store.
Rclone UI — Profile
The only entry that solves the trigger problem — 5,167 lines of Rust registering jobs with launchd, crontab, and Task Scheduler behind one trait.
restic — Profile
Its on-disk layout is data/<ab>/<sha256> plus snapshots/<id> — which is corpora-builder's unbuilt Phase 4, already shipped and hardened since 2014.
Seafile — Profile
Git's object model with content-defined chunking underneath and a human-readable description on every commit — and the S3 backend is the paid tier.
Syncthing — Profile
Closes a question two Lossless docs left open: a receive-only device does not silently orphan the collaborator's edit — it flags it, counts it, and offers Revert.
12-Factor Agents — Profile
Nothing to install — thirteen essays whose answer to "who owns the prompts, the context, the state, the control flow?" is always *you*.
5ire — Profile
5ire's MCP layer has no apply button — editing a server row in PGlite makes the client reconnect via the live-query changefeed.
Agent2Agent (A2A) Protocol — Profile
The load-bearing commitment isn't the JSON-RPC — it's *opacity*. An A2A agent publishes skills, never its prompts, memory, or tools.
AGENTS.md — Profile
The entire specification is one precedence rule: the closest `AGENTS.md` to the edited file wins. No schema, no validator.
Aider — Profile
Aider has no tool-call loop and no MCP client — not a thin one, none at all. It reads diffs straight out of plain chat text.
AnythingLLM — Profile
A "workspace" is a config row over one always-shared backend: the chat provider varies per workspace, the vector DB never can.
AutoGen — Profile
Every AutoGen team type overrides exactly one method — `select_speaker()` — and that single seam is the whole difference between them.
Beads — Profile
Not one embedding anywhere in the recall path — retrieval is a SQL view and a graph walk over the agent's own work graph.
ByteRover CLI — Profile
The memory store is a directory of Markdown you can cat, grep, and diff, with its own git history. Recall is BM25, not embeddings.
Chroma — Profile
The hybrid retrieval in Chroma's docs raises `NotImplementedError` on a laptop — `Collection.search()` is gated behind Cloud.
Cline — Profile
"Memory Bank", the cross-session memory pattern Cline is best known for, has zero lines of code behind it.
Codex — Profile
Codex compiles a real Apple Seatbelt profile and Landlock ruleset: a fully autonomous session that never prompts is still jailed by the OS.
Continue — Profile
Five differently-scoped assistants can ship side by side in one project and swap mid-session — hot-swapping a skillset per project.
Delta-Mem — Profile
Memory as a dense rank-8 state matrix inside each attention head, updated token by token — the model never looks anything up.
Dive — Profile
Two native shells, neither containing a line of agent logic — both are thin supervisors spawning the same Python subprocess.
Frictionless Specs — Profile
This one specs *datasets*: every RFC 2119 "MUST" in the prose maps onto a constraint in a published JSON-Schema profile.
GSD (Get Shit Done) — Profile
GSD's wager is that the limiting factor is context rot, not spec rigor — so it ships no spec grammar and resets the window per task.
Goose — Profile
In Goose, "extension" just means "a live MCP client" — and two variants skip the subprocess entirely to run in-process.
Graphify — Profile
The whole memory layer is one in-process NetworkX graph and a `graph.json` you can copy around — no database anywhere.
Graphiti — Profile
`expired_at` is when the engine retired a fact; `invalid_at` is when it stopped being true in the world. Routinely different dates.
Hindsight — Profile
`fact_type` is a database CHECK constraint: world facts and first-person experiences travel through different extraction prompts.
Honcho — Profile
Collections are unique on (observer, observed, workspace) — Honcho indexes memory by relationship rather than by user.
Kaas — Profile
Every credential, provider quirk, and even context assembly lives in Rust; React gets typed `invoke()` calls and nothing else.
LanceDB — Profile
The database *is* the file format: a `.lance` directory DuckDB, Polars, and Pandas read with no LanceDB process running.
Letta — Profile
The one entry where the agent itself — not an extraction pipeline, not a background summarizer — is the memory manager.
llms.txt — Profile
The only behavioral semantic in the whole format is an H2 named `## Optional` — links there may be dropped when context is tight.
MCP-API — Profile
However many APIs a user registers, the model only ever sees two tools: `ListAvailableApis` and `CallApi`.
MCP Python SDK — Profile
The pinned checkout is the unstable half of a rework: v2 swaps `FastMCP`/`ClientSession` out for `MCPServer`/`Client`.
Mem0 — Profile
Despite the graph-memory marketing, the graph is an array of linked memory IDs hanging off spaCy-extracted entities.
MemPalace — Profile
The counter-bet: extract nothing. Raw 800-character chunks score 96.6% R@5 on LongMemEval at zero LLM calls.
Milvus — Profile
One Milvus deployment is not a process — it is nine roles, three external dependencies, and a C++ ANN core reached through cgo.
Model Context Protocol (MCP) — Profile
The three *client* primitives — Sampling, Roots, Elicitation — let a server use the host's model without ever getting the user's key.
Neo — Profile
Familiar high-confidence queries get dialled down to low reasoning effort — memory shows up as a smaller bill, not better accuracy.
OpenSpec — Profile
An LLM should never re-emit a whole spec to change one line, so changes ship as `ADDED` / `MODIFIED` / `REMOVED` deltas.
Open Vibe — Profile
One real OS child process per workspace — but only ever one binary. What varies per workspace is the invocation, not the backend.
OpenAgent — Profile
Delegation recursion is prevented not by a depth counter but by never registering `spawn_agent` on the child's tool server.
OpenCode — Profile
A terminal coding agent with 38 SQLite migrations and real OTLP tracing — built like a production SaaS backend, and no Go anywhere.
OpenHands — Profile
Pin `OpenHands/OpenHands` today and you do not get the agent — the CodeAct loop moved out to a versioned pip dependency.
OpenUI — Profile
JSON spends ~30% of its bytes on punctuation an LLM has no reason to emit; the same seven UIs cost 47–67% fewer tokens in OpenUI Lang.
OpenViking — Profile
Memories, resources, and skills collapse into one `viking://` tree, and retrieval is a recursive directory walk, not a vector query.
Paxm — Profile
Paxm is not a memory store — it is routing and durability in front of seven-plus pluggable backends, and it does zero semantic dedup.
pgvector — Profile
About 3k lines of substantive C, and it inherits MVCC, WAL, replication, PITR, and row-level security completely untouched.
Pi — Profile
Pi's docs refuse MCP, sub-agents, permission popups, and plan mode by name — then ship every one as a first-party example extension.
Qdrant — Profile
Qdrant estimates a filter's cardinality *before* searching, which is why recall doesn't collapse the moment you filter on `tenant_id`.
RetainDB — Profile
Supersession as an actual indexed column — `validUntil`, `supersededBy` — rather than something the ranker is hoped to handle.
Routa — Profile
The swappable unit turns out to be the session, not the workspace: two sessions in one workspace can run two different agent CLIs.
Spec Kit — Profile
No plan gets past the Simplicity, Anti-Abstraction, and Integration-First gates without a written, paper-trailed exception.
StateBench — Profile
The trade nobody advertises: more context raises decision accuracy and superseded-fact resurrection together.
Supermemory — Profile
The most polished client surface in the study, wrapped around an engine that is not in the repo — every SDK points at a hosted API.
Superpowers — Profile
A SessionStart hook injects the bootstrap skill as `<EXTREMELY_IMPORTANT>` before the agent has written its first word.
Symphony — Profile
OpenAI shipped the spec, not the product: 2169 lines of RFC 2119 with sections *deliberately made redundant* for coding agents.
Understand-Anything — Profile
When an LLM authors your graph, the schema layer stops being a validator and becomes a repair shop — alias maps fold synonyms back.
Volt — Profile
Numeric thresholds decide when to compact, and the model is only ever asked to write summaries. Long-horizon coding as a DB problem.
Weaviate — Profile
The database itself calls OpenAI to embed on write and Cohere to rerank on read — the RAG pipeline lives inside the store, not the app.
agent-harnesses/profiles
conversational-ui-and-native-shells/profiles
memory-layers-for-agents/profiles
open-specs-and-standards/profiles
sync-and-content-version-control/profiles
vector-databases/profiles
floaters/7
inquiry/5
Agent Harnesses — Answers to the Plan's Targeted Questions
Phase 2, not a survey: three questions — MCP tool scoping, trajectory logs, per-project config swaps — answered from the profiles.
Conversational UI / Native Shells — Answers to the Plan's Targeted Questions
Phase 2, not a survey: two questions — Dive's Electron/Tauri migration and anything-llm's workspace switch — answered from the profiles.
agent-harnesses/inquiry
conversational-ui-and-native-shells/inquiry
open-specs-and-standards/inquiry