Skip to Content

Sessions, Context & Working Memory

Long-running interactions require intelligent context handling. In isagent.io/app, conversations are not just ephemeral chat logs; they are backed by structured session state, automatic context compaction, and persistent working memory.

Session Management

The left sidebar of the workspace allows users to organize their work streams:

  • Pinning & Tags: Pin critical ongoing investigations or tag sessions by project (e.g., #migration, #q4-audit).
  • Deep-Linked URLs: Every session is addressable at https://isagent.io/app/c/<sessionId>, making it simple to bookmark or share with teammates.
  • Session Export: Export complete conversations as structured Markdown, JSON-RPC transcripts, or standalone PDF summaries.

Context Window Compaction

When a conversation approaches the context window limit (e.g., 80% of model capacity after extensive file analysis or tool runs), the platform’s Context Compactor triggers automatically:

┌────────────────────────────────────────────────────────────────────────┐
│ Context Compactor Active:                                              │
│ • Initial Context: 142,000 tokens                                      │
│ • Summarized History: Core decisions & schema definitions preserved    │
│ • Compacted Context: 24,500 tokens (82% space recovered)              │
└────────────────────────────────────────────────────────────────────────┘
  1. Lossless Anchor Retention: File paths, code definitions, API keys (masked), and user constraints are permanently anchored in working memory.
  2. Intermediate Cleanup: Raw stdout logs and repetitive tool poll payloads are compressed into semantic summaries.
  3. Transparent Notification: A badge informs the user that context was optimized without losing task continuity.

File Attachments & Knowledge Grounding

Users can drag and drop external files directly into the prompt box:

  • Supported Formats: .pdf, .docx, .xlsx, .csv, .png, .jpg, .json, .sql, .py, .ts, .md.
  • RAG Vector Grounding: For files larger than 1 MB, the document is chunked and indexed in real-time, allowing the agent to retrieve exact sections via semantic similarity.
  • Multimodal Vision: Screenshots and UI wireframes are parsed natively by vision-capable models, enabling direct layout critique and React code generation.

Persistent Working Memory & Preferences

Under /settings, users can configure cross-session working memory:

  • Custom Instructions: Define global engineering standards (e.g., “Always write type-safe TypeScript with strict null checks” or “Format financial tables with currency symbols”).
  • Project Memories: Tell the agent about your organizational conventions, default branch names, or preferred deployment targets.