Coding agents are good at the code in front of them and bad at everything around it: why a table was denormalized last quarter, which caller actually binds to which overload, what the team decided about retry behaviour, and what the last session already tried. That gap is rarely a model intelligence problem. It is a context problem, and a large share of the agent tooling published in recent weeks is aimed squarely at closing it.
Code graphs instead of text search
The most direct fix is to stop asking agents to guess structure from filenames. Graphicle is a single Rust program that maps a repository into a code graph, remembers your project docs, decisions and past agent sessions, and returns cited excerpts with the exact files and lines. It runs on your machine with no API key, no account and no telemetry, and it works with Claude Code, Codex, Cursor and any MCP client. The citations are the useful part: an agent answer you can check against a line number is worth several you cannot.
For C# specifically, Graphify C# goes further than a graph and uses the compiler. It is a free, headless Roslyn and MSBuild indexer that turns a solution into deterministic semantic evidence: exact callers, references, implementations, inheritance, overrides and argument bindings, even across overloads, generics and project boundaries. Text search can find a spelling; it cannot tell you which overload a call binds to. This can, and it needs no IDE and no database.
Memory that survives the session
A code graph tells an agent where things are. It does not tell it what was decided. Friday is a self-hosted persistent cognitive memory layer for AI coding agents, built to stop Cursor, Claude and Copilot from having amnesia. It persists architecture decisions, schemas and constraints across sessions and exposes them over MCP against a knowledge graph.
Lossless Memory takes a deliberately opposite position for personal assistants: never summarize. Raw conversation logs are stored in full, every record carries a timestamp, and every index is built on that time axis, which the project calls the Temporal Backbone. You query by time first and words second, results come back chronological and unsummarized, and a small index tells the model which topic the conversation is in so context survives compaction. The tradeoff is storage growth, which is the honest price of not throwing anything away.
The same idea applied to a whole digital life rather than one assistant is Omnesis. It syncs from 40+ sources, including email, messages, notes, calendar, health and files, into a single local store, works out how people and documents connect, and exposes search, graph traversal and SQL over it behind a privacy gate that can allow, reduce, ask or deny each request. Everything stays on your own hardware.
Analysis that an agent can actually do
Data work fails differently from code work: the agent writes plausible SQL against the wrong join. Graphene is a data analytics framework built for coding agents, described as everything-as-code analytics. It gives agents two things: a semantic layer, where Graphene SQL combines SQL with governed metrics and modelled joins, and a dashboard file type that produces more consistent visuals than raw Python or JavaScript. Everything is driven from the CLI, and the languages are designed to be token-efficient.
Carrying artifacts and decisions between sessions
Context also leaks away in the handoff. Derive is review and approval for work made by AI agents: every version, comment, revision and decision lives at one durable URL, your team reviews the work, and any compatible agent can act on the feedback. It is MCP-native, fair source, and runs hosted or as a single self-hosted container.
Agentgit solves the narrower problem of passing work between two agents. It is a public git host where no account, token or key is required: you push to a name and the repository exists. The handoff becomes a URL that any sandbox can already receive, with no invite, no token and no archive of a working directory.
For output that is not code at all, Botbin is pastebin for AI agent artifacts. An agent posts a single-file HTML page with one curl line and gets a clean URL back, instead of dumping raw HTML into Telegram, Slack or a terminal. There is an MCP server with publish, update and get, so an agent can revise an artifact it already published.
Reducing what the agent has to hold in context
Two projects attack context from the supply side. Aclif is a framework for building command-line tools that agents use to reach SaaS platforms. One grammar and canonical names reach the same record by the same name on any provider, commands self-describe with schema, examples and safety metadata, and the same command runs from a shell, inside a tool call or behind a gateway. The motivation is explicit: an MCP server publishes a fixed tool list that occupies context on every turn, so publishing every operation is unaffordable while publishing a handful is incomplete.
Skillbay approaches the same problem as a marketplace. It is a curated catalogue of SKILL.md packages that teach a coding agent how to do a job, where every skill is reviewed by a person and shows the same prompt answered without and with the skill. The catalogue is available as a JSON API with an OpenAPI 3.1 spec, and an MCP server lets you browse and install free skills without signing in.
Where to start
If your agent keeps asking questions the repository already answers, start with Graphicle or Graphify C#. If it keeps re-deriving decisions, add Friday or Lossless Memory. If the failure is analysis rather than navigation, Graphene is the more targeted fix, and if the problem is that yesterday's work has no home, Derive, Agentgit and Botbin are the plumbing that gives it one.
