[Platform] Make memory.md context read cap configurable per agent (currently hardcoded to 2,000 chars)

[up-guillaume]

Problem

In agent_context.py, the memory.md content injected into agent context is truncated at 2,000 characters. This is hardcoded and applies uniformly to all agents.

2,000 characters is insufficient for agents with rich operational context:

• DevOps Moiria needs infrastructure mappings, container IDs, delegation rules
• Platform Moiria needs DB schema references, migration paths
• Curator needs the full 12-dimension monitoring framework

Agents are forced to choose between: (a) keeping memory.md under 2,000 chars (losing critical context), or (b) letting it grow and having the platform silently truncate it (agents don't know what context they're missing).

Root Cause

In the Clawith platform code (agent_context.py or similar), memory.md is read and truncated:

# Current (hardcoded)
memory_content = read_file(memory_path)[:2000]

The 2,000-char cap was likely set as a safe default to control token usage, but it's too aggressive for operational agents.

Proposed Solution

Make the cap configurable per agent, with a sane default:

# Proposed: per-agent configuration
DEFAULT_MEMORY_CAP = 5000  # chars

memory_cap = agent_config.get("memory_char_limit", DEFAULT_MEMORY_CAP)
memory_content = read_file(memory_path)[:memory_cap]

Configuration options:

1. Agent-level config (preferred): Add a memory_char_limit field to the agent profile in the database
2. Soul.md directive: Parse # memory_char_limit: 5000 from soul.md frontmatter
3. Environment variable: CLAWITH_MEMORY_CHAR_LIMIT as a global override

Recommended values:

Tier	Chars	Use case
Minimal	2,000	Simple agents with little context
Standard	5,000	Default for new agents
Extended	10,000	Operational agents (DevOps, Platform)
Maximum	20,000	Agents with large reference needs

Acceptance Criteria

• Identify the code path where memory.md is read and truncated in agent_context.py
• Add a configurable parameter (agent-level or global) with default of 5,000
• Log a warning when truncation occurs so agents know context was cut
• Document the configuration option
• Backward compatible: existing agents without the config field use the new default (5,000)

Context

• Discovered during Moiria Unified File System Cleanup (2026-08-02)
• Related: Issue [Platform] Deprecate CRITICAL_FILE_PATTERNS backup of write_file overwrites — causes backup file explosion #808 (CRITICAL_FILE_PATTERNS deprecation)
• Related: Issue Agent workspace indexing: hybrid roadmap (manifest.json → FTS5 → vector/RAG) #807 (hybrid workspace indexing roadmap)
• Related: Issue [Feature Request] Agent long-term memory: persist tool calling history and enable task resumption #454 (agent long-term memory persistence)

Priority

P2 — Agents are losing context silently. Not breaking, but degrading agent effectiveness.

Labels

enhancement, agent-context, configurability

[saisharan0103]

backend/app/services/agent_context.py is the part I would inspect first, since the reported behavior is a fixed 2,000-character slice of memory.md before injection. For a PR, I would start by threading a memory_char_limit from agent configuration with a default cap, keeping the current token-control behavior explicit. The regression test should cover the default path and an agent override above 2,000 chars, asserting the injected memory follows the configured limit.