The last day of open source AI on GitHub is not a list. It is a graph.

Two questions stacked into one. The first half 'open source AI projects releases GitHub last day' is a time-scoped activity query: everything with topic:artificial-intelligence (or topic:llm, topic:agents, topic:ml) that was pushed or tagged in the last 24 UTC hours. The second half '2026' is a year filter, and it matters because the ecosystem in April 2026 is dominated by agent protocols (ACP, MCP), local inference engines (llama.cpp, ollama, vllm, mlc-llm, mistral.rs), open-weight model families (DeepSeek-R1, GLM-5.1, Qwen3, Llama 4, Kimi-Dev-72B), and memory systems (MemPalace, a2a). A list of repos tells you what shipped. A graph of their relationships tells you why, and which nodes move together.

Because the interesting question is almost never 'which repo got the most commits,' it is 'which of these releases are related.' A vllm kernel tweak the same day DeepSeek-R1 drops a new checkpoint is one connected event, not two. A CrewAI security patch followed within hours by a downstream release in smolagents is a dependency edge, not a coincidence. GitHub's native search URL returns the nodes for free. The edges have to come from somewhere, which is what Fazm's save_knowledge_graph tool is for. The enum of node types is fixed at exactly five values (person, organization, place, thing, concept) so edges have something to attach to, and the live force-directed layout makes the clusters obvious at a glance.

Read them yourself at Desktop/Sources/MainWindow/Pages/MemoryGraph/ForceDirectedSimulation.swift lines 49 through 56. The defaults are repulsion 80,000, attraction 0.003, centerGravity 0.0008, damping 0.9, dt 0.016 (60 fps), restLength 600, maxSpeed 40. Nodes are SIMD3<Float> 3D vectors, seeded at a random position inside a sphere of radius 200 to 600, and the simulation considers itself stable when 10 consecutive frames fall below a 0.2 movement threshold. The constants retune themselves for small graphs: under a threshold in populate(), restLength drops to 300 and repulsion to 50,000, so a 12-node digest does not fly apart.

MemoryGraphPage.swift line 143 sets maxVisibleNodes = 100. That is the cap on what the force-directed view renders, not the cap on what the underlying KnowledgeGraphStorage holds. A typical last-24-hour pull on topic:artificial-intelligence returns on the order of 900 pushed repos and 60 tagged releases, so the agent is expected to rank and slice before it calls save_knowledge_graph. The tool description in fazm-tools-stdio.ts suggests 15 to 40 nodes with meaningful edges. That range is where the 3D layout looks its best.

MemoryGraphPage.swift line 299 has a Task.sleep of 5_000_000_000 nanoseconds wrapped inside addGraphFromStorage(). The reason: when a large digest lands, the agent may call save_knowledge_graph 4 or 5 times back-to-back (first the orgs, then the models, then the inference engines, then the agent frameworks, then the cross-edges). Without the debounce the graph view would re-layout after each call and the simulation would never stabilize. The buffer coalesces them into a single performIncrementalGraphUpdate() pass, which is the only thing that preserves the satisfying 'snap into place' feel.

The data comes from the same place the flat list does: the rendered GitHub search page. Fazm opens the search URL in Chrome, reads the result list via the macOS accessibility tree, extracts typed fields per repository card, and hands the rows to the model. The model is then prompted to emit entities (orgs, projects, models, concepts) and relationships (depends_on, forks, same_author, same_day_release) and call save_knowledge_graph with the shape. The raw repo rows can also be written to fazm.db via execute_sql for queryable history. The graph is a view over those two stores, not a separate pipeline.

Awesome-lists are human-maintained and lag by days or weeks. Monthly roundups are frozen the moment they publish. GitHub's topic pages sort by stars, so a 2023 repo that has not been touched in a year still sits above a repo that shipped a major release this morning. None of them answers the cross-project question: which projects in the ecosystem shipped together on the same day, and what do they have in common. A daily graph captures exactly that. As a consumer app, Fazm is also not a developer pipeline you have to maintain; install the DMG, grant Accessibility permission, and ask the agent to do it.

Semantically almost the same thing, but people who type 'last day 2026' are often asking one of three variations: the last 24 UTC hours, the last full calendar day in their timezone, or the final day of a named period like 'the last day of April 2026.' The GitHub URL pattern handles all three with different date inputs. Fazm resolves the intent in chat. If you say 'the last day,' it defaults to yesterday UTC. If you say 'the last day of April,' it builds a query for 2026-04-30. Either way the output is a graph with today's release activity clustered around stable anchor nodes.

No. Fazm is a consumer Mac app. You open the chat and say 'get the last day of open source AI releases from GitHub and show me the graph.' The agent handles the URL, the AX tree read, the entity extraction, the save_knowledge_graph call, and the 3D layout. The five node types are enforced by the JSON schema in fazm-tools-stdio.ts so the model cannot emit ambiguous types even if it wanted to. You watch the graph form in the MemoryGraphPage window while the simulation relaxes.

Yes. Repository is github.com/mediar-ai/fazm. License is MIT. The Swift desktop app lives under Desktop/, the Node ACP bridge under acp-bridge/. Every file this page cites (ForceDirectedSimulation.swift, MemoryGraphPage.swift, fazm-tools-stdio.ts, ChatPrompts.swift) is in that tree. Fazm itself shows up on your daily topic:artificial-intelligence pull whenever we push, which is a fun recursive property: the tool you use to visualize the ecosystem is a node inside the visualization.