Maintain Project Context Across Sessions

Context Agent is a session-continuity skill for Claude Code that captures, compresses and restores context automatically across sessions - topics, decisions, tasks, errors, modified files and technical discoveries. It is documented in Brazilian Portuguese, with only the When to Use and Limitations headers in English for catalog discovery and validation.

It writes context, logs, a SQLite database and MEMORY.md to local paths, and its own maintenance routine can archive and remove old summaries, so it carries an explicit warning to confirm paths, keep a backup, and get explicit approval before running init, save, archive or maintain.

Its CLI, context_manager.py, exposes six commands. init is run once, the first time the skill is used. save, run when a session is ending or before a long task, finds the most recent session JSONL file, parses every message, tool call and result, generates a structured session-NNN.md summary, updates ACTIVE_CONTEXT.md with new information, syncs it to MEMORY.md (which loads into every session's system prompt), and indexes everything for full-text search. load, run at the start of a new session, generates a briefing with active projects, pending tasks by priority, blockers, recent decisions, established conventions and a summary of recent sessions. status prints a short few-line summary of projects, critical pending items and blockers. search "<query>" runs a SQLite FTS5 full-text search across every session's topics, decisions, errors and files. maintain archives old sessions, compresses the archive, re-syncs MEMORY.md, and rebuilds the search index.

Each session file captures topics discussed, decisions with their rationale, completed tasks, pending tasks with priority, modified files, technical discoveries, resolved errors, open questions, technical debt, and metrics (input/output/cache tokens, messages, tool calls). ACTIVE_CONTEXT.md is capped at 150 lines and holds an active-projects table, pending tasks split by priority, recent decisions, active blockers, established conventions and a list of recent sessions.

The compression rules govern how that 150-line cap is maintained: a session becomes archive-eligible after being more than 20 sessions old (configurable via ARCHIVE_AFTER_SESSIONS); durable sections - topics, decisions, unfinished pending tasks, discoveries and resolved errors - are always kept, while ephemeral sections - metrics and granular modified-file detail - are dropped on archiving; once the archive accumulates 5 or more individual sessions they consolidate into a single ARCHIVE_YYYY.md keeping only headers and decisions; and ACTIVE_CONTEXT.md itself is pruned by removing completed tasks automatically on each save, aging out decisions after 30 days (configurable), keeping only the last 5 recent sessions, and dropping resolved blockers. A drift check compares ACTIVE_CONTEXT.md against MEMORY.md and, if they have diverged through manual editing or corruption, maintain resynchronizes them automatically.

Use it when the user mentions saving context, the next session, a session briefing, a session summary, or session continuity. Do not use it for a task unrelated to context continuity, when a simpler, more specific tool would serve the request, or when the user just needs general assistance without this specialization. The bundled implementation uses absolute paths specific to Windows and its original author, so every path must be adapted and validated before any execution, and its output is never a substitute for validation, testing or expert review.

Input is the current session's JSONL transcript (for save) or a search query (for search). Output is a session-NNN.md summary, an updated ACTIVE_CONTEXT.md and synced MEMORY.md, a PROJECT_REGISTRY.md status table, and, from load, a full briefing of active projects, prioritized pending tasks, blockers, recent decisions and conventions.

It integrates with Claude Code's session JSONL format, a SQLite FTS5 virtual table for full-text search, and MEMORY.md, which is included in every session's system prompt so a new Claude session starts already knowing project state without manual action. A companion skill, context-guardian, preserves context before context-window compaction.

It suits developers working across many Claude Code sessions on the same projects who need automatic continuity - a briefing of pending work, decisions and blockers at the start of each session - without manually re-explaining project state each time.