Aleph

/aleph - External Memory Workflow

TL;DR: Load large data into external memory, search it, reason in loops, and persist across sessions.

Quick Start

# Test if Aleph is available
list_contexts()

If that works, the MCP server is running.

Instant pattern:

load_context(content="<paste huge content here>", context_id="doc")
search_context(pattern="keyword", context_id="doc")
finalize(answer="Found X at line Y", context_id="doc")

Note: tool names may appear as mcp__aleph__load_context in your MCP client.

Instant RLM (Load File → Work)

The most common pattern: point at a file, let Aleph load it, immediately apply RLM reasoning.

load_file(path="path/to/large_file.md", context_id="doc")
search_context(pattern="relevant", context_id="doc")
exec_python(code="""
chunks = chunk(50000)
summaries = sub_query_batch("Summarize:", chunks)
print(summaries)
""", context_id="doc")
finalize(answer="...", context_id="doc")

When a user says /aleph myfile.py or $aleph myfile.py, load it and immediately begin this pattern.

Practical Defaults

• Use output="json" for structured results and output="markdown" for human-readable output.
• Action tools require starting the server with --enable-actions and may require confirm=true.
• For large docs, pair chunk_context() with peek_context() to navigate quickly.
• Use rg_search() for fast repo search and semantic_search() for meaning-based lookup.
• load_file() handles PDFs, Word docs, HTML, and compressed logs (.gz/.bz2/.xz).
• Save and resume long tasks with save_session() and load_session().
• Memory packs auto-save to .aleph/memory_pack.json and auto-load on startup (actions enabled).
• To see full tool docstrings, start with --tool-docs full or set ALEPH_TOOL_DOCS=full.

Core Patterns

1) Analyze Data

load_context(content=data_text, context_id="doc")
search_context(pattern="important|keyword|pattern", context_id="doc")
peek_context(start=100, end=150, unit="lines", context_id="doc")
finalize(answer="Analysis complete: ...", confidence="high", context_id="doc")

2) Compare Two Contexts

load_context(content=doc1, context_id="v1")
load_context(content=doc2, context_id="v2")
diff_contexts(a="v1", b="v2")
search_context(pattern="difference", context_id="v1")
search_context(pattern="difference", context_id="v2")
finalize(answer="Key differences: ...", context_id="v1")

3) Deep Reasoning Loop

load_context(content=problem, context_id="analysis")

think(question="What is the core issue?", context_id="analysis")
search_context(pattern="relevant", context_id="analysis")
evaluate_progress(
    current_understanding="I found X...",
    remaining_questions=["What about Y?"],
    confidence_score=0.7,
    context_id="analysis"
)

finalize(answer="Conclusion: ...", confidence="high", context_id="analysis")

4) Fast Repo Search (rg)

rg_search(pattern="TODO|FIXME", paths=["."], load_context_id="rg_hits", confirm=true)
search_context(pattern="TODO", context_id="rg_hits")

5) Semantic Search

semantic_search(query="login failure", context_id="doc", top_k=3)
peek_context(start=1200, end=1600, unit="chars", context_id="doc")

6) Recipe Pipelines (Declarative + DSL)

Use recipes when you want reusable, inspectable workflows.

A) Run JSON recipe directly

run_recipe(
  recipe={
    "version": "aleph.recipe.v1",
    "context_id": "doc",
    "budget": {"max_steps": 8, "max_sub_queries": 5},
    "steps": [
      {"op": "search", "pattern": "ERROR|WARN", "max_results": 10},
      {"op": "map_sub_query", "prompt": "Root cause?", "context_field": "context"},
      {"op": "aggregate", "prompt": "Synthesize causes"},
      {"op": "finalize"}
    ]
  }
)

B) Compile + run DSL code

run_recipe_code(
  context_id="doc",
  code="""
recipe = (
    Recipe(context_id='doc', max_sub_queries=5)
    .search('ERROR|WARN', max_results=10)
    .map_sub_query('Root cause?', context_field='context')
    .aggregate('Synthesize causes')
    .finalize()
)
"""
)

C) Chunk & Summarize (large docs)

run_recipe_code(
  context_id="doc",
  code="""
recipe = (
    Recipe(context_id='doc', max_sub_queries=5)
    .chunk(100000)
    .map_sub_query('Summarize this section')
    .aggregate('Combine into a unified summary')
    .finalize()
)
"""
)

D) Needle-in-Haystack (no sub-queries)

run_recipe_code(
  context_id="codebase",
  code="""
recipe = (
    Recipe(context_id='codebase')
    .search('TODO|FIXME|HACK|XXX', max_results=50)
    .filter(field='match', contains='HACK')
    .take(5)
    .finalize()
)
"""
)

Recommended flow: validate_recipe -> estimate_recipe -> run_recipe.

Ops: search, peek, lines, take, chunk, filter, map_sub_query, sub_query, aggregate, assign, load, finalize.

Sub-Query Guidance (RLM Best Practices)

Use sub-queries inside exec_python so the recursion is driven by code (symbolic loops), not by repeated tool calls. This follows the Recursive Language Model (RLM) paradigm.

Batching Efficiency (CRITICAL)

Sub-LLMs can handle ~500K characters per call. Batch aggressively to minimize API costs!

Rule of thumb: Aim for ~100-200K characters per sub_query call.

Example 1: Basic Chunking

exec_python(code=\"\"\"
chunks = chunk(100000)  # 100K char chunks
summaries = sub_query_batch(\"Summarize this chunk:\", chunks)
final = sub_query_strict(
    f\"Combine summaries into 5 bullets: {summaries}\",
    validate_regex=r\"^- \",
    max_retries=2,
)
print(final)
\"\"\", context_id=\"doc\")

Example 2: Iterative Document Analysis

For structured documents (books, papers, logs), iterate section-by-section:

exec_python(code=\"\"\"
query = "What caused the system failure?"
buffers = []

# Split by sections/headers
import re
sections = re.split(r'\\n## ', ctx)

for i, section in enumerate(sections):
    if len(section) < 500:  # skip tiny sections
        continue
    summary = sub_query(
        f"Extract info relevant to: {query}",
        context_slice=section
    )
    buffers.append(f"Section {i}: {summary}")
    print(f"Processed section {i}/{len(sections)}")

# Final aggregation
final = sub_query(
    f"Based on these findings, answer: {query}\\n\\n" + "\\n".join(buffers)
)
print(f"ANSWER: {final}")
\"\"\", context_id=\"doc\")

Example 3: Regex-Targeted Sub-Queries

When you know what to look for, use regex to narrow down, then sub-query on hits:

exec_python(code=\"\"\"
# Find relevant sections first
hits = search(r"error|exception|failed", max_results=20)
answers = []

for hit in hits:
    # Get surrounding context (100 lines)
    start = max(0, hit['line_num'] - 50)
    end = hit['line_num'] + 50
    snippet = lines(start, end)

    answer = sub_query(
        f"What error occurred here and what's the root cause?",
        context_slice=snippet
    )
    answers.append(f"Line {hit['line_num']}: {answer}")

# Aggregate findings
print("\\n".join(answers))
\"\"\", context_id=\"doc\")

Example 4: Answer Verification Pattern

Use sub-queries to verify answers and avoid context rot:

exec_python(code=\"\"\"
# First pass: extract candidate answer
candidate = sub_query("What is the magic number mentioned?", ctx[:200000])
print(f"Candidate: {candidate}")

# Verification pass: confirm with different context slice
verification = sub_query(
    f"Is '{candidate}' the correct magic number? Verify from this context.",
    ctx[200000:400000]
)
print(f"Verification: {verification}")
\"\"\", context_id=\"doc\")

If you need to parse results in exec_python, prefer line-based formats like KEY: value.

Sub-Query Backend Configuration

Sub-queries require a backend. Configure once per session:

Quick switch (REPL helper inside exec_python):

exec_python(code="set_backend('kimi')")  # or 'claude', 'codex', 'gemini', 'api'
exec_python(code="print(get_config())")    # verify current settings

MCP tool (direct call):

configure(sub_query_backend="kimi")
configure(sub_query_share_session=false)
configure(sub_query_timeout=300)
configure(sandbox_timeout=300)  # increase exec_python timeout for sub-query loops

Backend priority (auto mode): codex → gemini → kimi → claude → api

Per-call overrides: validate_regex and max_retries in sub_query_strict() override env defaults.

When a user says "use claude backend" or "switch to gemini", call set_backend() or configure().

Tool Reference

Core Tools (always available)

Context Management:

Search & Navigation:

Reasoning & Execution:

Action Tools (requires --enable-actions)

Filesystem:

Shell & Search:

Persistence:

Remote MCP Orchestration:

exec_python Helpers

Core:

• ctx, peek(start, end), lines(start, end), search(pattern), chunk(size)
• ctx_append(text) - append text to context (persists for subsequent operations)
• ctx_set(text) - replace entire context with new text
• semantic_search(query, ...) for meaning-based search
• embed_text(text, dim) for lightweight embeddings
• extract_routes(lang="auto") for route extraction
• cite(snippet, line_range, note) for evidence
• sub_query(prompt, context_slice) for recursion
• sub_aleph(query, context=None) for nested recursion
• sub_query_map(prompts, context_slices=None, limit=None) for batch sub-queries
• sub_query_batch(prompt, context_slices, limit=None) for one prompt over many slices
• sub_query_strict(prompt, context_slice=None, validate_regex=None, max_retries=0) for strict output validation
• Recipe DSL: Recipe(), Search(), Chunk(), Filter(), MapSubQuery(), Aggregate(), Finalize(), as_recipe()

100+ built-in helpers including:

• Extractors: extract_emails(), extract_urls(), extract_dates(), extract_ips(), extract_functions()
• Statistics: word_count(), line_count(), word_frequency(), ngrams()
• Line ops: head(), tail(), grep(), sort_lines(), columns()
• Validation: is_email(), is_url(), is_json(), is_numeric()

Extractors return list[dict] with keys: value, line_num, start, end.

Configure Reference

The configure() tool adjusts runtime settings without restarting:

Important: Set sub_query_share_session=false unless you specifically need sub-agents to access parent Aleph contexts. Session sharing adds complexity and can cause issues with some backends.

Troubleshooting

• "Tool not found": MCP server not running; check aleph command.
• "Context not found": verify context_id and use list_contexts().
• Search returns nothing: broaden your regex pattern.
• rg_search slow: install ripgrep (rg) for best performance.
• Running out of context: use summarize_so_far() to compress.
• Session file not found: check the path relative to the server working dir.